Merge branch 'master' into feature/deploy-environment-header

.ci/windows_amd_base_files/run_amd_gpu_enable_dynamic_vram.bat

@@ -1,2 +1,2 @@
-.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --disable-smart-memory
+.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --enable-dynamic-vram
 pause

.github/workflows/tag-dispatch-cloud.yml

@@ -0,0 +1,45 @@
+name: Tag Dispatch to Cloud
+
+on:
+  push:
+    tags:
+      - 'v*'
+
+jobs:
+  dispatch-cloud:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Send repository dispatch to cloud
+        env:
+          DISPATCH_TOKEN: ${{ secrets.CLOUD_REPO_DISPATCH_TOKEN }}
+          RELEASE_TAG: ${{ github.ref_name }}
+        run: |
+          set -euo pipefail
+
+          if [ -z "${DISPATCH_TOKEN:-}" ]; then
+            echo "::error::CLOUD_REPO_DISPATCH_TOKEN is required but not set."
+            exit 1
+          fi
+
+          RELEASE_URL="https://github.com/${{ github.repository }}/releases/tag/${RELEASE_TAG}"
+
+          PAYLOAD="$(jq -n \
+            --arg release_tag "$RELEASE_TAG" \
+            --arg release_url "$RELEASE_URL" \
+            '{
+              event_type: "comfyui_tag_pushed",
+              client_payload: {
+                release_tag: $release_tag,
+                release_url: $release_url
+              }
+            }')"
+
+          curl -fsSL \
+            -X POST \
+            -H "Accept: application/vnd.github+json" \
+            -H "Content-Type: application/json" \
+            -H "Authorization: Bearer ${DISPATCH_TOKEN}" \
+            https://api.github.com/repos/Comfy-Org/cloud/dispatches \
+            -d "$PAYLOAD"
+
+          echo "✅ Dispatched ComfyUI tag ${RELEASE_TAG} to Comfy-Org/cloud"

.gitignore

@@ -21,7 +21,6 @@ venv*/
 *.log
 web_custom_versions/
 .DS_Store
-openapi.yaml
 filtered-openapi.yaml
 uv.lock
 .comfy_environment

CODEOWNERS

@@ -1,2 +1,2 @@
 # Admins
-* @comfyanonymous @kosinkadink @guill
+* @comfyanonymous @kosinkadink @guill @alexisrolland @rattus128 @kijai

README.md

@@ -1,7 +1,7 @@
 <div align="center">
 
 # ComfyUI
-**The most powerful and modular visual AI engine and application.**
+**The most powerful and modular AI engine for content creation.**
 
 
 [![Website][website-shield]][website-url]
@@ -31,10 +31,16 @@
 [github-downloads-latest-shield]: https://img.shields.io/github/downloads/comfyanonymous/ComfyUI/latest/total?style=flat&label=downloads%40latest
 [github-downloads-link]: https://github.com/comfyanonymous/ComfyUI/releases
 
-![ComfyUI Screenshot](https://github.com/user-attachments/assets/7ccaf2c1-9b72-41ae-9a89-5688c94b7abe)
+<img width="1590" height="795" alt="ComfyUI Screenshot" src="https://github.com/user-attachments/assets/36e065e0-bfae-4456-8c7f-8369d5ea48a2" />
+<br>
 </div>
 
-ComfyUI lets you design and execute advanced stable diffusion pipelines using a graph/nodes/flowchart based interface. Available on Windows, Linux, and macOS.
+ComfyUI is the AI creation engine for visual professionals who demand control over every model, every parameter, and every output. Its powerful and modular node graph interface empowers creatives to generate images, videos, 3D models, audio, and more...
+- ComfyUI natively supports the latest open-source state of the art models.
+- API nodes provide access to the best closed source models such as Nano Banana, Seedance, Hunyuan3D, etc.
+- It is available on Windows, Linux, and macOS, locally with our desktop application or on our cloud.
+- The most sophisticated workflows can be exposed through a simple UI thanks to App Mode.
+- It integrates seamlessly into production pipelines with our API endpoints.
 
 ## Get Started
 
@@ -77,6 +83,7 @@ See what ComfyUI can do with the [newer template workflows](https://comfy.org/wo
    - [Hunyuan Image 2.1](https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_image/)
    - [Flux 2](https://comfyanonymous.github.io/ComfyUI_examples/flux2/)
    - [Z Image](https://comfyanonymous.github.io/ComfyUI_examples/z_image/)
+   - Ernie Image
 - Image Editing Models
    - [Omnigen 2](https://comfyanonymous.github.io/ComfyUI_examples/omnigen/)
    - [Flux Kontext](https://comfyanonymous.github.io/ComfyUI_examples/flux/#flux-kontext-image-editing-model)
@@ -193,13 +200,15 @@ If you have trouble extracting it, right click the file -> properties -> unblock
 
 The portable above currently comes with python 3.13 and pytorch cuda 13.0. Update your Nvidia drivers if it doesn't start.
 
-#### Alternative Downloads:
+#### All Official Portable Downloads:
 
 [Portable for AMD GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_amd.7z)
 
-[Experimental portable for Intel GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_intel.7z)
+[Portable for Intel GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_intel.7z)
 
-[Portable with pytorch cuda 12.6 and python 3.12](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia_cu126.7z) (Supports Nvidia 10 series and older GPUs).
+[Portable for Nvidia GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia.7z) (supports 20 series and above).
+
+[Portable for Nvidia GPUs with pytorch cuda 12.6 and python 3.12](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia_cu126.7z) (Supports Nvidia 10 series and older GPUs).
 
 #### How do I share models between another UI and ComfyUI?
 

blueprints/.glsl/Glow_30.frag

@@ -2,7 +2,6 @@
 precision mediump float;
 
 uniform sampler2D u_image0;
-uniform vec2 u_resolution;
 uniform int u_int0;      // Blend mode
 uniform int u_int1;      // Color tint
 uniform float u_float0;  // Intensity
@@ -75,7 +74,7 @@ void main() {
     float t0 = threshold - 0.15;
     float t1 = threshold + 0.15;
     
-    vec2 texelSize = 1.0 / u_resolution;
+    vec2 texelSize = 1.0 / vec2(textureSize(u_image0, 0));
     float radius2 = radius * radius;
     
     float sampleScale = clamp(radius * 0.75, 0.35, 1.0);

blueprints/.glsl/Image_Blur_1.frag

@@ -12,7 +12,6 @@ const int RADIAL_SAMPLES = 12;
 const float RADIAL_STRENGTH = 0.0003;
 
 uniform sampler2D u_image0;
-uniform vec2 u_resolution;
 uniform int u_int0;      // Blur type (BLUR_GAUSSIAN, BLUR_BOX, BLUR_RADIAL)
 uniform float u_float0;  // Blur radius/amount
 uniform int u_pass;      // Pass index (0 = horizontal, 1 = vertical)
@@ -25,7 +24,7 @@ float gaussian(float x, float sigma) {
 }
 
 void main() {
-    vec2 texelSize = 1.0 / u_resolution;
+    vec2 texelSize = 1.0 / vec2(textureSize(u_image0, 0));
     float radius = max(u_float0, 0.0);
 
     // Radial (angular) blur - single pass, doesn't use separable

blueprints/.glsl/Sharpen_23.frag

@@ -2,14 +2,13 @@
 precision highp float;
 
 uniform sampler2D u_image0;
-uniform vec2 u_resolution;
 uniform float u_float0;  // strength [0.0 – 2.0] typical: 0.3–1.0
 
 in vec2 v_texCoord;
 layout(location = 0) out vec4 fragColor0;
 
 void main() {
-    vec2 texel = 1.0 / u_resolution;
+    vec2 texel = 1.0 / vec2(textureSize(u_image0, 0));
     
     // Sample center and neighbors
     vec4 center = texture(u_image0, v_texCoord);

blueprints/.glsl/Unsharp_Mask_26.frag

@@ -2,7 +2,6 @@
 precision highp float;
 
 uniform sampler2D u_image0;
-uniform vec2 u_resolution;
 uniform float u_float0;  // amount    [0.0 - 3.0]  typical: 0.5-1.5
 uniform float u_float1;  // radius    [0.5 - 10.0] blur radius in pixels
 uniform float u_float2;  // threshold [0.0 - 0.1]  min difference to sharpen
@@ -19,7 +18,7 @@ float getLuminance(vec3 color) {
 }
 
 void main() {
-    vec2 texel = 1.0 / u_resolution;
+    vec2 texel = 1.0 / vec2(textureSize(u_image0, 0));
     float radius = max(u_float1, 0.5);
     float amount = u_float0;
     float threshold = u_float2;

blueprints/Depth to Image (Z-Image-Turbo).json

@@ -160,7 +160,7 @@
         },
         "revision": 0,
         "config": {},
-        "name": "local-Depth to Image (Z-Image-Turbo)",
+        "name": "Depth to Image (Z-Image-Turbo)",
         "inputNode": {
           "id": -10,
           "bounding": [
@@ -2482,4 +2482,4 @@
     "VHS_KeepIntermediate": true
   },
   "version": 0.4
-}
+}

blueprints/Depth to Video (ltx 2.0).json

@@ -261,7 +261,7 @@
         },
         "revision": 0,
         "config": {},
-        "name": "local-Depth to Video (LTX 2.0)",
+        "name": "Depth to Video (LTX 2.0)",
         "inputNode": {
           "id": -10,
           "bounding": [
@@ -5208,4 +5208,4 @@
     "workflowRendererVersion": "LG"
   },
   "version": 0.4
-}
+}

blueprints/Glow.json

@@ -268,7 +268,7 @@
               "Node name for S&R": "GLSLShader"
             },
             "widgets_values": [
-              "#version 300 es\nprecision mediump float;\n\nuniform sampler2D u_image0;\nuniform vec2 u_resolution;\nuniform int u_int0;      // Blend mode\nuniform int u_int1;      // Color tint\nuniform float u_float0;  // Intensity\nuniform float u_float1;  // Radius\nuniform float u_float2;  // Threshold\n\nin vec2 v_texCoord;\nout vec4 fragColor;\n\nconst int BLEND_ADD      = 0;\nconst int BLEND_SCREEN   = 1;\nconst int BLEND_SOFT     = 2;\nconst int BLEND_OVERLAY  = 3;\nconst int BLEND_LIGHTEN  = 4;\n\nconst float GOLDEN_ANGLE = 2.39996323;\nconst int MAX_SAMPLES = 48;\nconst vec3 LUMA = vec3(0.299, 0.587, 0.114);\n\nfloat hash(vec2 p) {\n    p = fract(p * vec2(123.34, 456.21));\n    p += dot(p, p + 45.32);\n    return fract(p.x * p.y);\n}\n\nvec3 hexToRgb(int h) {\n    return vec3(\n        float((h >> 16) & 255),\n        float((h >> 8) & 255),\n        float(h & 255)\n    ) * (1.0 / 255.0);\n}\n\nvec3 blend(vec3 base, vec3 glow, int mode) {\n    if (mode == BLEND_SCREEN) {\n        return 1.0 - (1.0 - base) * (1.0 - glow);\n    }\n    if (mode == BLEND_SOFT) {\n        return mix(\n            base - (1.0 - 2.0 * glow) * base * (1.0 - base),\n            base + (2.0 * glow - 1.0) * (sqrt(base) - base),\n            step(0.5, glow)\n        );\n    }\n    if (mode == BLEND_OVERLAY) {\n        return mix(\n            2.0 * base * glow,\n            1.0 - 2.0 * (1.0 - base) * (1.0 - glow),\n            step(0.5, base)\n        );\n    }\n    if (mode == BLEND_LIGHTEN) {\n        return max(base, glow);\n    }\n    return base + glow;\n}\n\nvoid main() {\n    vec4 original = texture(u_image0, v_texCoord);\n    \n    float intensity = u_float0 * 0.05;\n    float radius = u_float1 * u_float1 * 0.012;\n    \n    if (intensity < 0.001 || radius < 0.1) {\n        fragColor = original;\n        return;\n    }\n    \n    float threshold = 1.0 - u_float2 * 0.01;\n    float t0 = threshold - 0.15;\n    float t1 = threshold + 0.15;\n    \n    vec2 texelSize = 1.0 / u_resolution;\n    float radius2 = radius * radius;\n    \n    float sampleScale = clamp(radius * 0.75, 0.35, 1.0);\n    int samples = int(float(MAX_SAMPLES) * sampleScale);\n    \n    float noise = hash(gl_FragCoord.xy);\n    float angleOffset = noise * GOLDEN_ANGLE;\n    float radiusJitter = 0.85 + noise * 0.3;\n    \n    float ca = cos(GOLDEN_ANGLE);\n    float sa = sin(GOLDEN_ANGLE);\n    vec2 dir = vec2(cos(angleOffset), sin(angleOffset));\n    \n    vec3 glow = vec3(0.0);\n    float totalWeight = 0.0;\n    \n    // Center tap\n    float centerMask = smoothstep(t0, t1, dot(original.rgb, LUMA));\n    glow += original.rgb * centerMask * 2.0;\n    totalWeight += 2.0;\n    \n    for (int i = 1; i < MAX_SAMPLES; i++) {\n        if (i >= samples) break;\n        \n        float fi = float(i);\n        float dist = sqrt(fi / float(samples)) * radius * radiusJitter;\n        \n        vec2 offset = dir * dist * texelSize;\n        vec3 c = texture(u_image0, v_texCoord + offset).rgb;\n        float mask = smoothstep(t0, t1, dot(c, LUMA));\n        \n        float w = 1.0 - (dist * dist) / (radius2 * 1.5);\n        w = max(w, 0.0);\n        w *= w;\n        \n        glow += c * mask * w;\n        totalWeight += w;\n        \n        dir = vec2(\n            dir.x * ca - dir.y * sa,\n            dir.x * sa + dir.y * ca\n        );\n    }\n    \n    glow *= intensity / max(totalWeight, 0.001);\n    \n    if (u_int1 > 0) {\n        glow *= hexToRgb(u_int1);\n    }\n    \n    vec3 result = blend(original.rgb, glow, u_int0);\n    result += (noise - 0.5) * (1.0 / 255.0);\n    \n    fragColor = vec4(clamp(result, 0.0, 1.0), original.a);\n}",
+              "#version 300 es\nprecision mediump float;\n\nuniform sampler2D u_image0;\nuniform int u_int0;      // Blend mode\nuniform int u_int1;      // Color tint\nuniform float u_float0;  // Intensity\nuniform float u_float1;  // Radius\nuniform float u_float2;  // Threshold\n\nin vec2 v_texCoord;\nout vec4 fragColor;\n\nconst int BLEND_ADD      = 0;\nconst int BLEND_SCREEN   = 1;\nconst int BLEND_SOFT     = 2;\nconst int BLEND_OVERLAY  = 3;\nconst int BLEND_LIGHTEN  = 4;\n\nconst float GOLDEN_ANGLE = 2.39996323;\nconst int MAX_SAMPLES = 48;\nconst vec3 LUMA = vec3(0.299, 0.587, 0.114);\n\nfloat hash(vec2 p) {\n    p = fract(p * vec2(123.34, 456.21));\n    p += dot(p, p + 45.32);\n    return fract(p.x * p.y);\n}\n\nvec3 hexToRgb(int h) {\n    return vec3(\n        float((h >> 16) & 255),\n        float((h >> 8) & 255),\n        float(h & 255)\n    ) * (1.0 / 255.0);\n}\n\nvec3 blend(vec3 base, vec3 glow, int mode) {\n    if (mode == BLEND_SCREEN) {\n        return 1.0 - (1.0 - base) * (1.0 - glow);\n    }\n    if (mode == BLEND_SOFT) {\n        return mix(\n            base - (1.0 - 2.0 * glow) * base * (1.0 - base),\n            base + (2.0 * glow - 1.0) * (sqrt(base) - base),\n            step(0.5, glow)\n        );\n    }\n    if (mode == BLEND_OVERLAY) {\n        return mix(\n            2.0 * base * glow,\n            1.0 - 2.0 * (1.0 - base) * (1.0 - glow),\n            step(0.5, base)\n        );\n    }\n    if (mode == BLEND_LIGHTEN) {\n        return max(base, glow);\n    }\n    return base + glow;\n}\n\nvoid main() {\n    vec4 original = texture(u_image0, v_texCoord);\n    \n    float intensity = u_float0 * 0.05;\n    float radius = u_float1 * u_float1 * 0.012;\n    \n    if (intensity < 0.001 || radius < 0.1) {\n        fragColor = original;\n        return;\n    }\n    \n    float threshold = 1.0 - u_float2 * 0.01;\n    float t0 = threshold - 0.15;\n    float t1 = threshold + 0.15;\n    \n    vec2 texelSize = 1.0 / vec2(textureSize(u_image0, 0));\n    float radius2 = radius * radius;\n    \n    float sampleScale = clamp(radius * 0.75, 0.35, 1.0);\n    int samples = int(float(MAX_SAMPLES) * sampleScale);\n    \n    float noise = hash(gl_FragCoord.xy);\n    float angleOffset = noise * GOLDEN_ANGLE;\n    float radiusJitter = 0.85 + noise * 0.3;\n    \n    float ca = cos(GOLDEN_ANGLE);\n    float sa = sin(GOLDEN_ANGLE);\n    vec2 dir = vec2(cos(angleOffset), sin(angleOffset));\n    \n    vec3 glow = vec3(0.0);\n    float totalWeight = 0.0;\n    \n    // Center tap\n    float centerMask = smoothstep(t0, t1, dot(original.rgb, LUMA));\n    glow += original.rgb * centerMask * 2.0;\n    totalWeight += 2.0;\n    \n    for (int i = 1; i < MAX_SAMPLES; i++) {\n        if (i >= samples) break;\n        \n        float fi = float(i);\n        float dist = sqrt(fi / float(samples)) * radius * radiusJitter;\n        \n        vec2 offset = dir * dist * texelSize;\n        vec3 c = texture(u_image0, v_texCoord + offset).rgb;\n        float mask = smoothstep(t0, t1, dot(c, LUMA));\n        \n        float w = 1.0 - (dist * dist) / (radius2 * 1.5);\n        w = max(w, 0.0);\n        w *= w;\n        \n        glow += c * mask * w;\n        totalWeight += w;\n        \n        dir = vec2(\n            dir.x * ca - dir.y * sa,\n            dir.x * sa + dir.y * ca\n        );\n    }\n    \n    glow *= intensity / max(totalWeight, 0.001);\n    \n    if (u_int1 > 0) {\n        glow *= hexToRgb(u_int1);\n    }\n    \n    vec3 result = blend(original.rgb, glow, u_int0);\n    result += (noise - 0.5) * (1.0 / 255.0);\n    \n    fragColor = vec4(clamp(result, 0.0, 1.0), original.a);\n}",
               "from_input"
             ]
           },

blueprints/Image Blur.json

@@ -331,7 +331,7 @@
               "Node name for S&R": "GLSLShader"
             },
             "widgets_values": [
-              "#version 300 es\n#pragma passes 2\nprecision highp float;\n\n// Blur type constants\nconst int BLUR_GAUSSIAN = 0;\nconst int BLUR_BOX = 1;\nconst int BLUR_RADIAL = 2;\n\n// Radial blur config\nconst int RADIAL_SAMPLES = 12;\nconst float RADIAL_STRENGTH = 0.0003;\n\nuniform sampler2D u_image0;\nuniform vec2 u_resolution;\nuniform int u_int0;      // Blur type (BLUR_GAUSSIAN, BLUR_BOX, BLUR_RADIAL)\nuniform float u_float0;  // Blur radius/amount\nuniform int u_pass;      // Pass index (0 = horizontal, 1 = vertical)\n\nin vec2 v_texCoord;\nlayout(location = 0) out vec4 fragColor0;\n\nfloat gaussian(float x, float sigma) {\n    return exp(-(x * x) / (2.0 * sigma * sigma));\n}\n\nvoid main() {\n    vec2 texelSize = 1.0 / u_resolution;\n    float radius = max(u_float0, 0.0);\n\n    // Radial (angular) blur - single pass, doesn't use separable\n    if (u_int0 == BLUR_RADIAL) {\n        // Only execute on first pass\n        if (u_pass > 0) {\n            fragColor0 = texture(u_image0, v_texCoord);\n            return;\n        }\n\n        vec2 center = vec2(0.5);\n        vec2 dir = v_texCoord - center;\n        float dist = length(dir);\n\n        if (dist < 1e-4) {\n            fragColor0 = texture(u_image0, v_texCoord);\n            return;\n        }\n\n        vec4 sum = vec4(0.0);\n        float totalWeight = 0.0;\n        float angleStep = radius * RADIAL_STRENGTH;\n\n        dir /= dist;\n\n        float cosStep = cos(angleStep);\n        float sinStep = sin(angleStep);\n\n        float negAngle = -float(RADIAL_SAMPLES) * angleStep;\n        vec2 rotDir = vec2(\n            dir.x * cos(negAngle) - dir.y * sin(negAngle),\n            dir.x * sin(negAngle) + dir.y * cos(negAngle)\n        );\n\n        for (int i = -RADIAL_SAMPLES; i <= RADIAL_SAMPLES; i++) {\n            vec2 uv = center + rotDir * dist;\n            float w = 1.0 - abs(float(i)) / float(RADIAL_SAMPLES);\n            sum += texture(u_image0, uv) * w;\n            totalWeight += w;\n\n            rotDir = vec2(\n                rotDir.x * cosStep - rotDir.y * sinStep,\n                rotDir.x * sinStep + rotDir.y * cosStep\n            );\n        }\n\n        fragColor0 = sum / max(totalWeight, 0.001);\n        return;\n    }\n\n    // Separable Gaussian / Box blur\n    int samples = int(ceil(radius));\n\n    if (samples == 0) {\n        fragColor0 = texture(u_image0, v_texCoord);\n        return;\n    }\n\n    // Direction: pass 0 = horizontal, pass 1 = vertical\n    vec2 dir = (u_pass == 0) ? vec2(1.0, 0.0) : vec2(0.0, 1.0);\n\n    vec4 color = vec4(0.0);\n    float totalWeight = 0.0;\n    float sigma = radius / 2.0;\n\n    for (int i = -samples; i <= samples; i++) {\n        vec2 offset = dir * float(i) * texelSize;\n        vec4 sample_color = texture(u_image0, v_texCoord + offset);\n\n        float weight;\n        if (u_int0 == BLUR_GAUSSIAN) {\n            weight = gaussian(float(i), sigma);\n        } else {\n            // BLUR_BOX\n            weight = 1.0;\n        }\n\n        color += sample_color * weight;\n        totalWeight += weight;\n    }\n\n    fragColor0 = color / totalWeight;\n}\n",
+              "#version 300 es\n#pragma passes 2\nprecision highp float;\n\n// Blur type constants\nconst int BLUR_GAUSSIAN = 0;\nconst int BLUR_BOX = 1;\nconst int BLUR_RADIAL = 2;\n\n// Radial blur config\nconst int RADIAL_SAMPLES = 12;\nconst float RADIAL_STRENGTH = 0.0003;\n\nuniform sampler2D u_image0;\nuniform int u_int0;      // Blur type (BLUR_GAUSSIAN, BLUR_BOX, BLUR_RADIAL)\nuniform float u_float0;  // Blur radius/amount\nuniform int u_pass;      // Pass index (0 = horizontal, 1 = vertical)\n\nin vec2 v_texCoord;\nlayout(location = 0) out vec4 fragColor0;\n\nfloat gaussian(float x, float sigma) {\n    return exp(-(x * x) / (2.0 * sigma * sigma));\n}\n\nvoid main() {\n    vec2 texelSize = 1.0 / vec2(textureSize(u_image0, 0));\n    float radius = max(u_float0, 0.0);\n\n    // Radial (angular) blur - single pass, doesn't use separable\n    if (u_int0 == BLUR_RADIAL) {\n        // Only execute on first pass\n        if (u_pass > 0) {\n            fragColor0 = texture(u_image0, v_texCoord);\n            return;\n        }\n\n        vec2 center = vec2(0.5);\n        vec2 dir = v_texCoord - center;\n        float dist = length(dir);\n\n        if (dist < 1e-4) {\n            fragColor0 = texture(u_image0, v_texCoord);\n            return;\n        }\n\n        vec4 sum = vec4(0.0);\n        float totalWeight = 0.0;\n        float angleStep = radius * RADIAL_STRENGTH;\n\n        dir /= dist;\n\n        float cosStep = cos(angleStep);\n        float sinStep = sin(angleStep);\n\n        float negAngle = -float(RADIAL_SAMPLES) * angleStep;\n        vec2 rotDir = vec2(\n            dir.x * cos(negAngle) - dir.y * sin(negAngle),\n            dir.x * sin(negAngle) + dir.y * cos(negAngle)\n        );\n\n        for (int i = -RADIAL_SAMPLES; i <= RADIAL_SAMPLES; i++) {\n            vec2 uv = center + rotDir * dist;\n            float w = 1.0 - abs(float(i)) / float(RADIAL_SAMPLES);\n            sum += texture(u_image0, uv) * w;\n            totalWeight += w;\n\n            rotDir = vec2(\n                rotDir.x * cosStep - rotDir.y * sinStep,\n                rotDir.x * sinStep + rotDir.y * cosStep\n            );\n        }\n\n        fragColor0 = sum / max(totalWeight, 0.001);\n        return;\n    }\n\n    // Separable Gaussian / Box blur\n    int samples = int(ceil(radius));\n\n    if (samples == 0) {\n        fragColor0 = texture(u_image0, v_texCoord);\n        return;\n    }\n\n    // Direction: pass 0 = horizontal, pass 1 = vertical\n    vec2 dir = (u_pass == 0) ? vec2(1.0, 0.0) : vec2(0.0, 1.0);\n\n    vec4 color = vec4(0.0);\n    float totalWeight = 0.0;\n    float sigma = radius / 2.0;\n\n    for (int i = -samples; i <= samples; i++) {\n        vec2 offset = dir * float(i) * texelSize;\n        vec4 sample_color = texture(u_image0, v_texCoord + offset);\n\n        float weight;\n        if (u_int0 == BLUR_GAUSSIAN) {\n            weight = gaussian(float(i), sigma);\n        } else {\n            // BLUR_BOX\n            weight = 1.0;\n        }\n\n        color += sample_color * weight;\n        totalWeight += weight;\n    }\n\n    fragColor0 = color / totalWeight;\n}\n",
               "from_input"
             ]
           }

blueprints/Image Edit (Flux.2 Klein 4B).json

@@ -128,7 +128,7 @@
         },
         "revision": 0,
         "config": {},
-        "name": "local-Image Edit (Flux.2 Klein 4B)",
+        "name": "Image Edit (Flux.2 Klein 4B)",
         "inputNode": {
           "id": -10,
           "bounding": [
@@ -1837,4 +1837,4 @@
     }
   },
   "version": 0.4
-}
+}

blueprints/Image Inpainting (Qwen-image).json

@@ -124,7 +124,7 @@
         },
         "revision": 0,
         "config": {},
-        "name": "local-Image Inpainting (Qwen-image)",
+        "name": "Image Inpainting (Qwen-image)",
         "inputNode": {
           "id": -10,
           "bounding": [
@@ -1923,4 +1923,4 @@
     "workflowRendererVersion": "LG"
   },
   "version": 0.4
-}
+}

blueprints/Image Outpainting (Qwen-Image).json

@@ -204,7 +204,7 @@
         },
         "revision": 0,
         "config": {},
-        "name": "local-Image Outpainting (Qwen-Image)",
+        "name": "Image Outpainting (Qwen-Image)",
         "inputNode": {
           "id": -10,
           "bounding": [
@@ -2749,4 +2749,4 @@
     }
   },
   "version": 0.4
-}
+}

blueprints/Image to Layers(Qwen-Image-Layered).json

@@ -1,15 +1,14 @@
 {
-  "id": "1a761372-7c82-4016-b9bf-fa285967e1e9",
   "revision": 0,
-  "last_node_id": 83,
+  "last_node_id": 176,
   "last_link_id": 0,
   "nodes": [
     {
-      "id": 83,
-      "type": "f754a936-daaf-4b6e-9658-41fdc54d301d",
+      "id": 176,
+      "type": "2d2e3c8e-53b3-4618-be52-6d1d99382f0e",
       "pos": [
-        61.999827823554256,
-        153.3332507624185
+        -1150,
+        200
       ],
       "size": [
         400,
@@ -56,6 +55,38 @@
             "name": "layers"
           },
           "link": null
+        },
+        {
+          "name": "seed",
+          "type": "INT",
+          "widget": {
+            "name": "seed"
+          },
+          "link": null
+        },
+        {
+          "name": "unet_name",
+          "type": "COMBO",
+          "widget": {
+            "name": "unet_name"
+          },
+          "link": null
+        },
+        {
+          "name": "clip_name",
+          "type": "COMBO",
+          "widget": {
+            "name": "clip_name"
+          },
+          "link": null
+        },
+        {
+          "name": "vae_name",
+          "type": "COMBO",
+          "widget": {
+            "name": "vae_name"
+          },
+          "link": null
         }
       ],
       "outputs": [
@@ -66,28 +97,41 @@
           "links": []
         }
       ],
+      "title": "Image to Layers (Qwen-Image-Layered)",
       "properties": {
         "proxyWidgets": [
           [
-            "-1",
+            "6",
             "text"
           ],
           [
-            "-1",
+            "3",
             "steps"
           ],
           [
-            "-1",
+            "3",
             "cfg"
           ],
           [
-            "-1",
+            "83",
             "layers"
           ],
           [
             "3",
             "seed"
           ],
+          [
+            "37",
+            "unet_name"
+          ],
+          [
+            "38",
+            "clip_name"
+          ],
+          [
+            "39",
+            "vae_name"
+          ],
           [
             "3",
             "control_after_generate"
@@ -95,6 +139,11 @@
         ],
         "cnr_id": "comfy-core",
         "ver": "0.5.1",
+        "ue_properties": {
+          "widget_ue_connectable": {},
+          "input_ue_unconnectable": {},
+          "version": "7.7"
+        },
         "enableTabs": false,
         "tabWidth": 65,
         "tabXOffset": 10,
@@ -103,25 +152,20 @@
         "secondTabOffset": 80,
         "secondTabWidth": 65
       },
-      "widgets_values": [
-        "",
-        20,
-        2.5,
-        2
-      ]
+      "widgets_values": []
     }
   ],
   "links": [],
-  "groups": [],
+  "version": 0.4,
   "definitions": {
     "subgraphs": [
       {
-        "id": "f754a936-daaf-4b6e-9658-41fdc54d301d",
+        "id": "2d2e3c8e-53b3-4618-be52-6d1d99382f0e",
         "version": 1,
         "state": {
-          "lastGroupId": 3,
-          "lastNodeId": 83,
-          "lastLinkId": 159,
+          "lastGroupId": 8,
+          "lastNodeId": 176,
+          "lastLinkId": 380,
           "lastRerouteId": 0
         },
         "revision": 0,
@@ -130,10 +174,10 @@
         "inputNode": {
           "id": -10,
           "bounding": [
-            -510,
-            523,
+            -720,
+            720,
             120,
-            140
+            220
           ]
         },
         "outputNode": {
@@ -156,8 +200,8 @@
             ],
             "localized_name": "image",
             "pos": [
-              -410,
-              543
+              -620,
+              740
             ]
           },
           {
@@ -168,8 +212,8 @@
               150
             ],
             "pos": [
-              -410,
-              563
+              -620,
+              760
             ]
           },
           {
@@ -180,8 +224,8 @@
               153
             ],
             "pos": [
-              -410,
-              583
+              -620,
+              780
             ]
           },
           {
@@ -192,8 +236,8 @@
               154
             ],
             "pos": [
-              -410,
-              603
+              -620,
+              800
             ]
           },
           {
@@ -204,8 +248,56 @@
               159
             ],
             "pos": [
-              -410,
-              623
+              -620,
+              820
+            ]
+          },
+          {
+            "id": "9f76338b-f4ca-4bb3-b61a-57b3f233061e",
+            "name": "seed",
+            "type": "INT",
+            "linkIds": [
+              377
+            ],
+            "pos": [
+              -620,
+              840
+            ]
+          },
+          {
+            "id": "8d0422d5-5eee-4f7e-9817-dc613cc62eca",
+            "name": "unet_name",
+            "type": "COMBO",
+            "linkIds": [
+              378
+            ],
+            "pos": [
+              -620,
+              860
+            ]
+          },
+          {
+            "id": "552eece2-a735-4d00-ae78-ded454622bc1",
+            "name": "clip_name",
+            "type": "COMBO",
+            "linkIds": [
+              379
+            ],
+            "pos": [
+              -620,
+              880
+            ]
+          },
+          {
+            "id": "1e6d141c-d0f9-4a2b-895c-b6780e57cfa0",
+            "name": "vae_name",
+            "type": "COMBO",
+            "linkIds": [
+              380
+            ],
+            "pos": [
+              -620,
+              900
             ]
           }
         ],
@@ -231,14 +323,14 @@
             "type": "CLIPLoader",
             "pos": [
               -320,
-              310
+              360
             ],
             "size": [
-              346.7470703125,
-              106
+              350,
+              150
             ],
             "flags": {},
-            "order": 0,
+            "order": 5,
             "mode": 0,
             "inputs": [
               {
@@ -248,7 +340,7 @@
                 "widget": {
                   "name": "clip_name"
                 },
-                "link": null
+                "link": 379
               },
               {
                 "localized_name": "type",
@@ -283,9 +375,14 @@
               }
             ],
             "properties": {
-              "Node name for S&R": "CLIPLoader",
               "cnr_id": "comfy-core",
               "ver": "0.5.1",
+              "ue_properties": {
+                "widget_ue_connectable": {},
+                "input_ue_unconnectable": {},
+                "version": "7.7"
+              },
+              "Node name for S&R": "CLIPLoader",
               "models": [
                 {
                   "name": "qwen_2.5_vl_7b_fp8_scaled.safetensors",
@@ -312,14 +409,14 @@
             "type": "VAELoader",
             "pos": [
               -320,
-              460
+              580
             ],
             "size": [
-              346.7470703125,
-              58
+              350,
+              110
             ],
             "flags": {},
-            "order": 1,
+            "order": 6,
             "mode": 0,
             "inputs": [
               {
@@ -329,7 +426,7 @@
                 "widget": {
                   "name": "vae_name"
                 },
-                "link": null
+                "link": 380
               }
             ],
             "outputs": [
@@ -345,9 +442,14 @@
               }
             ],
             "properties": {
-              "Node name for S&R": "VAELoader",
               "cnr_id": "comfy-core",
               "ver": "0.5.1",
+              "ue_properties": {
+                "widget_ue_connectable": {},
+                "input_ue_unconnectable": {},
+                "version": "7.7"
+              },
+              "Node name for S&R": "VAELoader",
               "models": [
                 {
                   "name": "qwen_image_layered_vae.safetensors",
@@ -375,11 +477,11 @@
               420
             ],
             "size": [
-              425.27801513671875,
-              180.6060791015625
+              430,
+              190
             ],
             "flags": {},
-            "order": 3,
+            "order": 2,
             "mode": 0,
             "inputs": [
               {
@@ -411,9 +513,14 @@
             ],
             "title": "CLIP Text Encode (Negative Prompt)",
             "properties": {
-              "Node name for S&R": "CLIPTextEncode",
               "cnr_id": "comfy-core",
               "ver": "0.5.1",
+              "ue_properties": {
+                "widget_ue_connectable": {},
+                "input_ue_unconnectable": {},
+                "version": "7.7"
+              },
+              "Node name for S&R": "CLIPTextEncode",
               "enableTabs": false,
               "tabWidth": 65,
               "tabXOffset": 10,
@@ -432,12 +539,12 @@
             "id": 70,
             "type": "ReferenceLatent",
             "pos": [
-              330,
-              670
+              140,
+              700
             ],
             "size": [
-              204.1666717529297,
-              46
+              210,
+              50
             ],
             "flags": {
               "collapsed": true
@@ -470,9 +577,14 @@
               }
             ],
             "properties": {
-              "Node name for S&R": "ReferenceLatent",
               "cnr_id": "comfy-core",
               "ver": "0.5.1",
+              "ue_properties": {
+                "widget_ue_connectable": {},
+                "input_ue_unconnectable": {},
+                "version": "7.7"
+              },
+              "Node name for S&R": "ReferenceLatent",
               "enableTabs": false,
               "tabWidth": 65,
               "tabXOffset": 10,
@@ -480,19 +592,18 @@
               "secondTabText": "Send Back",
               "secondTabOffset": 80,
               "secondTabWidth": 65
-            },
-            "widgets_values": []
+            }
           },
           {
             "id": 69,
             "type": "ReferenceLatent",
             "pos": [
-              330,
-              710
+              160,
+              820
             ],
             "size": [
-              204.1666717529297,
-              46
+              210,
+              50
             ],
             "flags": {
               "collapsed": true
@@ -525,9 +636,14 @@
               }
             ],
             "properties": {
-              "Node name for S&R": "ReferenceLatent",
               "cnr_id": "comfy-core",
               "ver": "0.5.1",
+              "ue_properties": {
+                "widget_ue_connectable": {},
+                "input_ue_unconnectable": {},
+                "version": "7.7"
+              },
+              "Node name for S&R": "ReferenceLatent",
               "enableTabs": false,
               "tabWidth": 65,
               "tabXOffset": 10,
@@ -535,8 +651,7 @@
               "secondTabText": "Send Back",
               "secondTabOffset": 80,
               "secondTabWidth": 65
-            },
-            "widgets_values": []
+            }
           },
           {
             "id": 66,
@@ -547,10 +662,10 @@
             ],
             "size": [
               270,
-              58
+              110
             ],
             "flags": {},
-            "order": 4,
+            "order": 7,
             "mode": 0,
             "inputs": [
               {
@@ -580,9 +695,14 @@
               }
             ],
             "properties": {
-              "Node name for S&R": "ModelSamplingAuraFlow",
               "cnr_id": "comfy-core",
               "ver": "0.5.1",
+              "ue_properties": {
+                "widget_ue_connectable": {},
+                "input_ue_unconnectable": {},
+                "version": "7.7"
+              },
+              "Node name for S&R": "ModelSamplingAuraFlow",
               "enableTabs": false,
               "tabWidth": 65,
               "tabXOffset": 10,
@@ -600,11 +720,11 @@
             "type": "LatentCutToBatch",
             "pos": [
               830,
-              160
+              140
             ],
             "size": [
               270,
-              82
+              140
             ],
             "flags": {},
             "order": 11,
@@ -646,9 +766,14 @@
               }
             ],
             "properties": {
-              "Node name for S&R": "LatentCutToBatch",
               "cnr_id": "comfy-core",
               "ver": "0.5.1",
+              "ue_properties": {
+                "widget_ue_connectable": {},
+                "input_ue_unconnectable": {},
+                "version": "7.7"
+              },
+              "Node name for S&R": "LatentCutToBatch",
               "enableTabs": false,
               "tabWidth": 65,
               "tabXOffset": 10,
@@ -666,12 +791,12 @@
             "id": 71,
             "type": "VAEEncode",
             "pos": [
-              100,
-              690
+              -280,
+              780
             ],
             "size": [
-              140,
-              46
+              230,
+              100
             ],
             "flags": {
               "collapsed": false
@@ -704,9 +829,14 @@
               }
             ],
             "properties": {
-              "Node name for S&R": "VAEEncode",
               "cnr_id": "comfy-core",
               "ver": "0.5.1",
+              "ue_properties": {
+                "widget_ue_connectable": {},
+                "input_ue_unconnectable": {},
+                "version": "7.7"
+              },
+              "Node name for S&R": "VAEEncode",
               "enableTabs": false,
               "tabWidth": 65,
               "tabXOffset": 10,
@@ -714,24 +844,23 @@
               "secondTabText": "Send Back",
               "secondTabOffset": 80,
               "secondTabWidth": 65
-            },
-            "widgets_values": []
+            }
           },
           {
             "id": 8,
             "type": "VAEDecode",
             "pos": [
               850,
-              310
+              370
             ],
             "size": [
               210,
-              46
+              50
             ],
             "flags": {
               "collapsed": true
             },
-            "order": 7,
+            "order": 3,
             "mode": 0,
             "inputs": [
               {
@@ -759,9 +888,14 @@
               }
             ],
             "properties": {
-              "Node name for S&R": "VAEDecode",
               "cnr_id": "comfy-core",
               "ver": "0.5.1",
+              "ue_properties": {
+                "widget_ue_connectable": {},
+                "input_ue_unconnectable": {},
+                "version": "7.7"
+              },
+              "Node name for S&R": "VAEDecode",
               "enableTabs": false,
               "tabWidth": 65,
               "tabXOffset": 10,
@@ -769,8 +903,7 @@
               "secondTabText": "Send Back",
               "secondTabOffset": 80,
               "secondTabWidth": 65
-            },
-            "widgets_values": []
+            }
           },
           {
             "id": 6,
@@ -780,11 +913,11 @@
               180
             ],
             "size": [
-              422.84503173828125,
-              164.31304931640625
+              430,
+              170
             ],
             "flags": {},
-            "order": 6,
+            "order": 1,
             "mode": 0,
             "inputs": [
               {
@@ -816,9 +949,14 @@
             ],
             "title": "CLIP Text Encode (Positive Prompt)",
             "properties": {
-              "Node name for S&R": "CLIPTextEncode",
               "cnr_id": "comfy-core",
               "ver": "0.5.1",
+              "ue_properties": {
+                "widget_ue_connectable": {},
+                "input_ue_unconnectable": {},
+                "version": "7.7"
+              },
+              "Node name for S&R": "CLIPTextEncode",
               "enableTabs": false,
               "tabWidth": 65,
               "tabXOffset": 10,
@@ -838,14 +976,14 @@
             "type": "KSampler",
             "pos": [
               530,
-              280
+              340
             ],
             "size": [
               270,
               400
             ],
             "flags": {},
-            "order": 5,
+            "order": 0,
             "mode": 0,
             "inputs": [
               {
@@ -879,7 +1017,7 @@
                 "widget": {
                   "name": "seed"
                 },
-                "link": null
+                "link": 377
               },
               {
                 "localized_name": "steps",
@@ -939,9 +1077,14 @@
               }
             ],
             "properties": {
-              "Node name for S&R": "KSampler",
               "cnr_id": "comfy-core",
               "ver": "0.5.1",
+              "ue_properties": {
+                "widget_ue_connectable": {},
+                "input_ue_unconnectable": {},
+                "version": "7.7"
+              },
+              "Node name for S&R": "KSampler",
               "enableTabs": false,
               "tabWidth": 65,
               "tabXOffset": 10,
@@ -964,12 +1107,12 @@
             "id": 78,
             "type": "GetImageSize",
             "pos": [
-              80,
-              790
+              -280,
+              930
             ],
             "size": [
-              210,
-              136
+              230,
+              140
             ],
             "flags": {},
             "order": 12,
@@ -1007,9 +1150,14 @@
               }
             ],
             "properties": {
-              "Node name for S&R": "GetImageSize",
               "cnr_id": "comfy-core",
               "ver": "0.5.1",
+              "ue_properties": {
+                "widget_ue_connectable": {},
+                "input_ue_unconnectable": {},
+                "version": "7.7"
+              },
+              "Node name for S&R": "GetImageSize",
               "enableTabs": false,
               "tabWidth": 65,
               "tabXOffset": 10,
@@ -1017,23 +1165,23 @@
               "secondTabText": "Send Back",
               "secondTabOffset": 80,
               "secondTabWidth": 65
-            },
-            "widgets_values": []
+            }
           },
           {
             "id": 83,
             "type": "EmptyQwenImageLayeredLatentImage",
             "pos": [
-              320,
-              790
+              -280,
+              1120
             ],
             "size": [
-              330.9341796875,
-              130
+              340,
+              200
             ],
             "flags": {},
             "order": 13,
             "mode": 0,
+            "showAdvanced": true,
             "inputs": [
               {
                 "localized_name": "width",
@@ -1083,9 +1231,14 @@
               }
             ],
             "properties": {
-              "Node name for S&R": "EmptyQwenImageLayeredLatentImage",
               "cnr_id": "comfy-core",
               "ver": "0.5.1",
+              "ue_properties": {
+                "widget_ue_connectable": {},
+                "input_ue_unconnectable": {},
+                "version": "7.7"
+              },
+              "Node name for S&R": "EmptyQwenImageLayeredLatentImage",
               "enableTabs": false,
               "tabWidth": 65,
               "tabXOffset": 10,
@@ -1109,11 +1262,11 @@
               180
             ],
             "size": [
-              346.7470703125,
-              82
+              350,
+              110
             ],
             "flags": {},
-            "order": 2,
+            "order": 4,
             "mode": 0,
             "inputs": [
               {
@@ -1123,7 +1276,7 @@
                 "widget": {
                   "name": "unet_name"
                 },
-                "link": null
+                "link": 378
               },
               {
                 "localized_name": "weight_dtype",
@@ -1147,9 +1300,14 @@
               }
             ],
             "properties": {
-              "Node name for S&R": "UNETLoader",
               "cnr_id": "comfy-core",
               "ver": "0.5.1",
+              "ue_properties": {
+                "widget_ue_connectable": {},
+                "input_ue_unconnectable": {},
+                "version": "7.7"
+              },
+              "Node name for S&R": "UNETLoader",
               "models": [
                 {
                   "name": "qwen_image_layered_bf16.safetensors",
@@ -1191,8 +1349,8 @@
             "bounding": [
               -330,
               110,
-              366.7470703125,
-              421.6
+              370,
+              610
             ],
             "color": "#3f789e",
             "font_size": 24,
@@ -1391,6 +1549,38 @@
             "target_id": 83,
             "target_slot": 2,
             "type": "INT"
+          },
+          {
+            "id": 377,
+            "origin_id": -10,
+            "origin_slot": 5,
+            "target_id": 3,
+            "target_slot": 4,
+            "type": "INT"
+          },
+          {
+            "id": 378,
+            "origin_id": -10,
+            "origin_slot": 6,
+            "target_id": 37,
+            "target_slot": 0,
+            "type": "COMBO"
+          },
+          {
+            "id": 379,
+            "origin_id": -10,
+            "origin_slot": 7,
+            "target_id": 38,
+            "target_slot": 0,
+            "type": "COMBO"
+          },
+          {
+            "id": 380,
+            "origin_id": -10,
+            "origin_slot": 8,
+            "target_id": 39,
+            "target_slot": 0,
+            "type": "COMBO"
           }
         ],
         "extra": {
@@ -1400,7 +1590,6 @@
       }
     ]
   },
-  "config": {},
   "extra": {
     "ds": {
       "scale": 1.14,
@@ -1409,7 +1598,6 @@
         6.855893974423647
       ]
     },
-    "workflowRendererVersion": "LG"
-  },
-  "version": 0.4
-}
+    "ue_links": []
+  }
+}

blueprints/Sharpen.json

@@ -267,7 +267,7 @@
               "Node name for S&R": "GLSLShader"
             },
             "widgets_values": [
-              "#version 300 es\nprecision highp float;\n\nuniform sampler2D u_image0;\nuniform vec2 u_resolution;\nuniform float u_float0;  // strength [0.0 – 2.0] typical: 0.3–1.0\n\nin vec2 v_texCoord;\nlayout(location = 0) out vec4 fragColor0;\n\nvoid main() {\n    vec2 texel = 1.0 / u_resolution;\n    \n    // Sample center and neighbors\n    vec4 center = texture(u_image0, v_texCoord);\n    vec4 top    = texture(u_image0, v_texCoord + vec2( 0.0, -texel.y));\n    vec4 bottom = texture(u_image0, v_texCoord + vec2( 0.0,  texel.y));\n    vec4 left   = texture(u_image0, v_texCoord + vec2(-texel.x,  0.0));\n    vec4 right  = texture(u_image0, v_texCoord + vec2( texel.x,  0.0));\n    \n    // Edge enhancement (Laplacian)\n    vec4 edges = center * 4.0 - top - bottom - left - right;\n    \n    // Add edges back scaled by strength\n    vec4 sharpened = center + edges * u_float0;\n    \n    fragColor0 = vec4(clamp(sharpened.rgb, 0.0, 1.0), center.a);\n}",
+              "#version 300 es\nprecision highp float;\n\nuniform sampler2D u_image0;\nuniform float u_float0;  // strength [0.0 – 2.0] typical: 0.3–1.0\n\nin vec2 v_texCoord;\nlayout(location = 0) out vec4 fragColor0;\n\nvoid main() {\n    vec2 texel = 1.0 / vec2(textureSize(u_image0, 0));\n    \n    // Sample center and neighbors\n    vec4 center = texture(u_image0, v_texCoord);\n    vec4 top    = texture(u_image0, v_texCoord + vec2( 0.0, -texel.y));\n    vec4 bottom = texture(u_image0, v_texCoord + vec2( 0.0,  texel.y));\n    vec4 left   = texture(u_image0, v_texCoord + vec2(-texel.x,  0.0));\n    vec4 right  = texture(u_image0, v_texCoord + vec2( texel.x,  0.0));\n    \n    // Edge enhancement (Laplacian)\n    vec4 edges = center * 4.0 - top - bottom - left - right;\n    \n    // Add edges back scaled by strength\n    vec4 sharpened = center + edges * u_float0;\n    \n    fragColor0 = vec4(clamp(sharpened.rgb, 0.0, 1.0), center.a);\n}",
               "from_input"
             ]
           }

blueprints/Unsharp Mask.json

@@ -383,7 +383,7 @@
               "Node name for S&R": "GLSLShader"
             },
             "widgets_values": [
-              "#version 300 es\nprecision highp float;\n\nuniform sampler2D u_image0;\nuniform vec2 u_resolution;\nuniform float u_float0;  // amount    [0.0 - 3.0]  typical: 0.5-1.5\nuniform float u_float1;  // radius    [0.5 - 10.0] blur radius in pixels\nuniform float u_float2;  // threshold [0.0 - 0.1]  min difference to sharpen\n\nin vec2 v_texCoord;\nlayout(location = 0) out vec4 fragColor0;\n\nfloat gaussian(float x, float sigma) {\n    return exp(-(x * x) / (2.0 * sigma * sigma));\n}\n\nfloat getLuminance(vec3 color) {\n    return dot(color, vec3(0.2126, 0.7152, 0.0722));\n}\n\nvoid main() {\n    vec2 texel = 1.0 / u_resolution;\n    float radius = max(u_float1, 0.5);\n    float amount = u_float0;\n    float threshold = u_float2;\n\n    vec4 original = texture(u_image0, v_texCoord);\n\n    // Gaussian blur for the \"unsharp\" mask\n    int samples = int(ceil(radius));\n    float sigma = radius / 2.0;\n\n    vec4 blurred = vec4(0.0);\n    float totalWeight = 0.0;\n\n    for (int x = -samples; x <= samples; x++) {\n        for (int y = -samples; y <= samples; y++) {\n            vec2 offset = vec2(float(x), float(y)) * texel;\n            vec4 sample_color = texture(u_image0, v_texCoord + offset);\n\n            float dist = length(vec2(float(x), float(y)));\n            float weight = gaussian(dist, sigma);\n            blurred += sample_color * weight;\n            totalWeight += weight;\n        }\n    }\n    blurred /= totalWeight;\n\n    // Unsharp mask = original - blurred\n    vec3 mask = original.rgb - blurred.rgb;\n\n    // Luminance-based threshold with smooth falloff\n    float lumaDelta = abs(getLuminance(original.rgb) - getLuminance(blurred.rgb));\n    float thresholdScale = smoothstep(0.0, threshold, lumaDelta);\n    mask *= thresholdScale;\n\n    // Sharpen: original + mask * amount\n    vec3 sharpened = original.rgb + mask * amount;\n\n    fragColor0 = vec4(clamp(sharpened, 0.0, 1.0), original.a);\n}\n",
+              "#version 300 es\nprecision highp float;\n\nuniform sampler2D u_image0;\nuniform float u_float0;  // amount    [0.0 - 3.0]  typical: 0.5-1.5\nuniform float u_float1;  // radius    [0.5 - 10.0] blur radius in pixels\nuniform float u_float2;  // threshold [0.0 - 0.1]  min difference to sharpen\n\nin vec2 v_texCoord;\nlayout(location = 0) out vec4 fragColor0;\n\nfloat gaussian(float x, float sigma) {\n    return exp(-(x * x) / (2.0 * sigma * sigma));\n}\n\nfloat getLuminance(vec3 color) {\n    return dot(color, vec3(0.2126, 0.7152, 0.0722));\n}\n\nvoid main() {\n    vec2 texel = 1.0 / vec2(textureSize(u_image0, 0));\n    float radius = max(u_float1, 0.5);\n    float amount = u_float0;\n    float threshold = u_float2;\n\n    vec4 original = texture(u_image0, v_texCoord);\n\n    // Gaussian blur for the \"unsharp\" mask\n    int samples = int(ceil(radius));\n    float sigma = radius / 2.0;\n\n    vec4 blurred = vec4(0.0);\n    float totalWeight = 0.0;\n\n    for (int x = -samples; x <= samples; x++) {\n        for (int y = -samples; y <= samples; y++) {\n            vec2 offset = vec2(float(x), float(y)) * texel;\n            vec4 sample_color = texture(u_image0, v_texCoord + offset);\n\n            float dist = length(vec2(float(x), float(y)));\n            float weight = gaussian(dist, sigma);\n            blurred += sample_color * weight;\n            totalWeight += weight;\n        }\n    }\n    blurred /= totalWeight;\n\n    // Unsharp mask = original - blurred\n    vec3 mask = original.rgb - blurred.rgb;\n\n    // Luminance-based threshold with smooth falloff\n    float lumaDelta = abs(getLuminance(original.rgb) - getLuminance(blurred.rgb));\n    float thresholdScale = smoothstep(0.0, threshold, lumaDelta);\n    mask *= thresholdScale;\n\n    // Sharpen: original + mask * amount\n    vec3 sharpened = original.rgb + mask * amount;\n\n    fragColor0 = vec4(clamp(sharpened, 0.0, 1.0), original.a);\n}\n",
               "from_input"
             ]
           }

comfy/cli_args.py

@@ -90,8 +90,8 @@ parser.add_argument("--force-channels-last", action="store_true", help="Force ch
 parser.add_argument("--directml", type=int, nargs="?", metavar="DIRECTML_DEVICE", const=-1, help="Use torch-directml.")
 
 parser.add_argument("--oneapi-device-selector", type=str, default=None, metavar="SELECTOR_STRING", help="Sets the oneAPI device(s) this instance will use.")
-parser.add_argument("--disable-ipex-optimize", action="store_true", help="Disables ipex.optimize default when loading models with Intel's Extension for Pytorch.")
 parser.add_argument("--supports-fp8-compute", action="store_true", help="ComfyUI will act like if the device supports fp8 compute.")
+parser.add_argument("--enable-triton-backend", action="store_true", help="ComfyUI will enable the use of Triton backend in comfy-kitchen. Is disabled at launch by default.")
 
 class LatentPreviewMethod(enum.Enum):
     NoPreviews = "none"

comfy/latent_formats.py

@@ -224,6 +224,7 @@ class Flux2(LatentFormat):
 
         self.latent_rgb_factors_bias = [-0.0329, -0.0718, -0.0851]
         self.latent_rgb_factors_reshape = lambda t: t.reshape(t.shape[0], 32, 2, 2, t.shape[-2], t.shape[-1]).permute(0, 1, 4, 2, 5, 3).reshape(t.shape[0], 32, t.shape[-2] * 2, t.shape[-1] * 2)
+        self.taesd_decoder_name = "taef2_decoder"
 
     def process_in(self, latent):
         return latent
@@ -783,3 +784,10 @@ class ZImagePixelSpace(ChromaRadiance):
     No VAE encoding/decoding — the model operates directly on RGB pixels.
     """
     pass
+
+class CogVideoX(LatentFormat):
+    latent_channels = 16
+    latent_dimensions = 3
+
+    def __init__(self):
+        self.scale_factor = 1.15258426

comfy/ldm/cogvideo/model.py

@@ -0,0 +1,573 @@
+# CogVideoX 3D Transformer - ported to ComfyUI native ops
+# Architecture reference: diffusers CogVideoXTransformer3DModel
+# Style reference: comfy/ldm/wan/model.py
+
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from comfy.ldm.modules.attention import optimized_attention
+import comfy.patcher_extension
+import comfy.ldm.common_dit
+
+
+def _get_1d_rotary_pos_embed(dim, pos, theta=10000.0):
+    """Returns (cos, sin) each with shape [seq_len, dim].
+
+    Frequencies are computed at dim//2 resolution then repeat_interleaved
+    to full dim, matching CogVideoX's interleaved (real, imag) pair format.
+    """
+    freqs = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float32, device=pos.device) / dim))
+    angles = torch.outer(pos.float(), freqs.float())
+    cos = angles.cos().repeat_interleave(2, dim=-1).float()
+    sin = angles.sin().repeat_interleave(2, dim=-1).float()
+    return (cos, sin)
+
+
+def apply_rotary_emb(x, freqs_cos_sin):
+    """Apply CogVideoX rotary embedding to query or key tensor.
+
+    x: [B, heads, seq_len, head_dim]
+    freqs_cos_sin: (cos, sin) each [seq_len, head_dim//2]
+
+    Uses interleaved pair rotation (same as diffusers CogVideoX/Flux).
+    head_dim is reshaped to (-1, 2) pairs, rotated, then flattened back.
+    """
+    cos, sin = freqs_cos_sin
+    cos = cos[None, None, :, :].to(x.device)
+    sin = sin[None, None, :, :].to(x.device)
+
+    # Interleaved pairs: [B, H, S, D] -> [B, H, S, D//2, 2] -> (real, imag)
+    x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1)
+    x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3)
+
+    return (x.float() * cos + x_rotated.float() * sin).to(x.dtype)
+
+
+def get_timestep_embedding(timesteps, dim, flip_sin_to_cos=True, downscale_freq_shift=0, scale=1, max_period=10000):
+    half = dim // 2
+    freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32, device=timesteps.device) / half)
+    args = timesteps[:, None].float() * freqs[None] * scale
+    embedding = torch.cat([torch.sin(args), torch.cos(args)], dim=-1)
+    if flip_sin_to_cos:
+        embedding = torch.cat([embedding[:, half:], embedding[:, :half]], dim=-1)
+    if dim % 2:
+        embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    return embedding
+
+
+def get_3d_sincos_pos_embed(embed_dim, spatial_size, temporal_size, spatial_interpolation_scale=1.0, temporal_interpolation_scale=1.0, device=None):
+    if isinstance(spatial_size, int):
+        spatial_size = (spatial_size, spatial_size)
+
+    grid_w = torch.arange(spatial_size[0], dtype=torch.float32, device=device) / spatial_interpolation_scale
+    grid_h = torch.arange(spatial_size[1], dtype=torch.float32, device=device) / spatial_interpolation_scale
+    grid_t = torch.arange(temporal_size, dtype=torch.float32, device=device) / temporal_interpolation_scale
+
+    grid_t, grid_h, grid_w = torch.meshgrid(grid_t, grid_h, grid_w, indexing="ij")
+
+    embed_dim_spatial = 2 * (embed_dim // 3)
+    embed_dim_temporal = embed_dim // 3
+
+    pos_embed_spatial = _get_2d_sincos_pos_embed(embed_dim_spatial, grid_h, grid_w, device=device)
+    pos_embed_temporal = _get_1d_sincos_pos_embed(embed_dim_temporal, grid_t[:, 0, 0], device=device)
+
+    T, H, W = grid_t.shape
+    pos_embed_temporal = pos_embed_temporal.unsqueeze(1).unsqueeze(1).expand(-1, H, W, -1)
+    pos_embed = torch.cat([pos_embed_temporal, pos_embed_spatial], dim=-1)
+
+    return pos_embed
+
+
+def _get_2d_sincos_pos_embed(embed_dim, grid_h, grid_w, device=None):
+    T, H, W = grid_h.shape
+    half_dim = embed_dim // 2
+    pos_h = _get_1d_sincos_pos_embed(half_dim, grid_h.reshape(-1), device=device).reshape(T, H, W, half_dim)
+    pos_w = _get_1d_sincos_pos_embed(half_dim, grid_w.reshape(-1), device=device).reshape(T, H, W, half_dim)
+    return torch.cat([pos_h, pos_w], dim=-1)
+
+
+def _get_1d_sincos_pos_embed(embed_dim, pos, device=None):
+    half = embed_dim // 2
+    freqs = torch.exp(-math.log(10000.0) * torch.arange(start=0, end=half, dtype=torch.float32, device=device) / half)
+    args = pos.float().reshape(-1)[:, None] * freqs[None]
+    embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+    if embed_dim % 2:
+        embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    return embedding
+
+
+
+class CogVideoXPatchEmbed(nn.Module):
+    def __init__(self, patch_size=2, patch_size_t=None, in_channels=16, dim=1920,
+                 text_dim=4096, bias=True, sample_width=90, sample_height=60,
+                 sample_frames=49, temporal_compression_ratio=4,
+                 max_text_seq_length=226, spatial_interpolation_scale=1.875,
+                 temporal_interpolation_scale=1.0, use_positional_embeddings=True,
+                 use_learned_positional_embeddings=True,
+                 device=None, dtype=None, operations=None):
+        super().__init__()
+        self.patch_size = patch_size
+        self.patch_size_t = patch_size_t
+        self.dim = dim
+        self.sample_height = sample_height
+        self.sample_width = sample_width
+        self.sample_frames = sample_frames
+        self.temporal_compression_ratio = temporal_compression_ratio
+        self.max_text_seq_length = max_text_seq_length
+        self.spatial_interpolation_scale = spatial_interpolation_scale
+        self.temporal_interpolation_scale = temporal_interpolation_scale
+        self.use_positional_embeddings = use_positional_embeddings
+        self.use_learned_positional_embeddings = use_learned_positional_embeddings
+
+        if patch_size_t is None:
+            self.proj = operations.Conv2d(in_channels, dim, kernel_size=patch_size, stride=patch_size, bias=bias, device=device, dtype=dtype)
+        else:
+            self.proj = operations.Linear(in_channels * patch_size * patch_size * patch_size_t, dim, device=device, dtype=dtype)
+
+        self.text_proj = operations.Linear(text_dim, dim, device=device, dtype=dtype)
+
+        if use_positional_embeddings or use_learned_positional_embeddings:
+            persistent = use_learned_positional_embeddings
+            pos_embedding = self._get_positional_embeddings(sample_height, sample_width, sample_frames)
+            self.register_buffer("pos_embedding", pos_embedding, persistent=persistent)
+
+    def _get_positional_embeddings(self, sample_height, sample_width, sample_frames, device=None):
+        post_patch_height = sample_height // self.patch_size
+        post_patch_width = sample_width // self.patch_size
+        post_time_compression_frames = (sample_frames - 1) // self.temporal_compression_ratio + 1
+        if self.patch_size_t is not None:
+            post_time_compression_frames = post_time_compression_frames // self.patch_size_t
+        num_patches = post_patch_height * post_patch_width * post_time_compression_frames
+
+        pos_embedding = get_3d_sincos_pos_embed(
+            self.dim,
+            (post_patch_width, post_patch_height),
+            post_time_compression_frames,
+            self.spatial_interpolation_scale,
+            self.temporal_interpolation_scale,
+            device=device,
+        )
+        pos_embedding = pos_embedding.reshape(-1, self.dim)
+        joint_pos_embedding = pos_embedding.new_zeros(
+            1, self.max_text_seq_length + num_patches, self.dim, requires_grad=False
+        )
+        joint_pos_embedding.data[:, self.max_text_seq_length:].copy_(pos_embedding)
+        return joint_pos_embedding
+
+    def forward(self, text_embeds, image_embeds):
+        input_dtype = text_embeds.dtype
+        text_embeds = self.text_proj(text_embeds.to(self.text_proj.weight.dtype)).to(input_dtype)
+        batch_size, num_frames, channels, height, width = image_embeds.shape
+
+        proj_dtype = self.proj.weight.dtype
+        if self.patch_size_t is None:
+            image_embeds = image_embeds.reshape(-1, channels, height, width)
+            image_embeds = self.proj(image_embeds.to(proj_dtype)).to(input_dtype)
+            image_embeds = image_embeds.view(batch_size, num_frames, *image_embeds.shape[1:])
+            image_embeds = image_embeds.flatten(3).transpose(2, 3)
+            image_embeds = image_embeds.flatten(1, 2)
+        else:
+            p = self.patch_size
+            p_t = self.patch_size_t
+            image_embeds = image_embeds.permute(0, 1, 3, 4, 2)
+            image_embeds = image_embeds.reshape(
+                batch_size, num_frames // p_t, p_t, height // p, p, width // p, p, channels
+            )
+            image_embeds = image_embeds.permute(0, 1, 3, 5, 7, 2, 4, 6).flatten(4, 7).flatten(1, 3)
+            image_embeds = self.proj(image_embeds.to(proj_dtype)).to(input_dtype)
+
+        embeds = torch.cat([text_embeds, image_embeds], dim=1).contiguous()
+
+        if self.use_positional_embeddings or self.use_learned_positional_embeddings:
+            text_seq_length = text_embeds.shape[1]
+            num_image_patches = image_embeds.shape[1]
+
+            if self.use_learned_positional_embeddings:
+                image_pos = self.pos_embedding[
+                    :, self.max_text_seq_length:self.max_text_seq_length + num_image_patches
+                ].to(device=embeds.device, dtype=embeds.dtype)
+            else:
+                image_pos = get_3d_sincos_pos_embed(
+                    self.dim,
+                    (width // self.patch_size, height // self.patch_size),
+                    num_image_patches // ((height // self.patch_size) * (width // self.patch_size)),
+                    self.spatial_interpolation_scale,
+                    self.temporal_interpolation_scale,
+                    device=embeds.device,
+                ).reshape(1, num_image_patches, self.dim).to(dtype=embeds.dtype)
+
+            # Build joint: zeros for text + sincos for image
+            joint_pos = torch.zeros(1, text_seq_length + num_image_patches, self.dim, device=embeds.device, dtype=embeds.dtype)
+            joint_pos[:, text_seq_length:] = image_pos
+            embeds = embeds + joint_pos
+
+        return embeds
+
+
+class CogVideoXLayerNormZero(nn.Module):
+    def __init__(self, time_dim, dim, elementwise_affine=True, eps=1e-5, bias=True,
+                 device=None, dtype=None, operations=None):
+        super().__init__()
+        self.silu = nn.SiLU()
+        self.linear = operations.Linear(time_dim, 6 * dim, bias=bias, device=device, dtype=dtype)
+        self.norm = operations.LayerNorm(dim, eps=eps, elementwise_affine=elementwise_affine, device=device, dtype=dtype)
+
+    def forward(self, hidden_states, encoder_hidden_states, temb):
+        shift, scale, gate, enc_shift, enc_scale, enc_gate = self.linear(self.silu(temb)).chunk(6, dim=1)
+        hidden_states = self.norm(hidden_states) * (1 + scale)[:, None, :] + shift[:, None, :]
+        encoder_hidden_states = self.norm(encoder_hidden_states) * (1 + enc_scale)[:, None, :] + enc_shift[:, None, :]
+        return hidden_states, encoder_hidden_states, gate[:, None, :], enc_gate[:, None, :]
+
+
+class CogVideoXAdaLayerNorm(nn.Module):
+    def __init__(self, time_dim, dim, elementwise_affine=True, eps=1e-5,
+                 device=None, dtype=None, operations=None):
+        super().__init__()
+        self.silu = nn.SiLU()
+        self.linear = operations.Linear(time_dim, 2 * dim, device=device, dtype=dtype)
+        self.norm = operations.LayerNorm(dim, eps=eps, elementwise_affine=elementwise_affine, device=device, dtype=dtype)
+
+    def forward(self, x, temb):
+        temb = self.linear(self.silu(temb))
+        shift, scale = temb.chunk(2, dim=1)
+        x = self.norm(x) * (1 + scale)[:, None, :] + shift[:, None, :]
+        return x
+
+
+class CogVideoXBlock(nn.Module):
+    def __init__(self, dim, num_heads, head_dim, time_dim,
+                 eps=1e-5, ff_inner_dim=None, ff_bias=True,
+                 device=None, dtype=None, operations=None):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = head_dim
+
+        self.norm1 = CogVideoXLayerNormZero(time_dim, dim, eps=eps, device=device, dtype=dtype, operations=operations)
+
+        # Self-attention (joint text + latent)
+        self.q = operations.Linear(dim, dim, bias=True, device=device, dtype=dtype)
+        self.k = operations.Linear(dim, dim, bias=True, device=device, dtype=dtype)
+        self.v = operations.Linear(dim, dim, bias=True, device=device, dtype=dtype)
+        self.norm_q = operations.LayerNorm(head_dim, eps=1e-6, elementwise_affine=True, device=device, dtype=dtype)
+        self.norm_k = operations.LayerNorm(head_dim, eps=1e-6, elementwise_affine=True, device=device, dtype=dtype)
+        self.attn_out = operations.Linear(dim, dim, bias=True, device=device, dtype=dtype)
+
+        self.norm2 = CogVideoXLayerNormZero(time_dim, dim, eps=eps, device=device, dtype=dtype, operations=operations)
+
+        # Feed-forward (GELU approximate)
+        inner_dim = ff_inner_dim or dim * 4
+        self.ff_proj = operations.Linear(dim, inner_dim, bias=ff_bias, device=device, dtype=dtype)
+        self.ff_out = operations.Linear(inner_dim, dim, bias=ff_bias, device=device, dtype=dtype)
+
+    def forward(self, hidden_states, encoder_hidden_states, temb, image_rotary_emb=None, transformer_options=None):
+        if transformer_options is None:
+            transformer_options = {}
+        text_seq_length = encoder_hidden_states.size(1)
+
+        # Norm & modulate
+        norm_hidden, norm_encoder, gate_msa, enc_gate_msa = self.norm1(hidden_states, encoder_hidden_states, temb)
+
+        # Joint self-attention
+        qkv_input = torch.cat([norm_encoder, norm_hidden], dim=1)
+        b, s, _ = qkv_input.shape
+        n, d = self.num_heads, self.head_dim
+
+        q = self.q(qkv_input).view(b, s, n, d)
+        k = self.k(qkv_input).view(b, s, n, d)
+        v = self.v(qkv_input)
+
+        q = self.norm_q(q).view(b, s, n, d)
+        k = self.norm_k(k).view(b, s, n, d)
+
+        # Apply rotary embeddings to image tokens only (diffusers format: [B, heads, seq, head_dim])
+        if image_rotary_emb is not None:
+            q_img = q[:, text_seq_length:].transpose(1, 2)  # [B, heads, img_seq, head_dim]
+            k_img = k[:, text_seq_length:].transpose(1, 2)
+            q_img = apply_rotary_emb(q_img, image_rotary_emb)
+            k_img = apply_rotary_emb(k_img, image_rotary_emb)
+            q = torch.cat([q[:, :text_seq_length], q_img.transpose(1, 2)], dim=1)
+            k = torch.cat([k[:, :text_seq_length], k_img.transpose(1, 2)], dim=1)
+
+        attn_out = optimized_attention(
+            q.reshape(b, s, n * d),
+            k.reshape(b, s, n * d),
+            v,
+            heads=self.num_heads,
+            transformer_options=transformer_options,
+        )
+
+        attn_out = self.attn_out(attn_out)
+
+        attn_encoder, attn_hidden = attn_out.split([text_seq_length, s - text_seq_length], dim=1)
+
+        hidden_states = hidden_states + gate_msa * attn_hidden
+        encoder_hidden_states = encoder_hidden_states + enc_gate_msa * attn_encoder
+
+        # Norm & modulate for FF
+        norm_hidden, norm_encoder, gate_ff, enc_gate_ff = self.norm2(hidden_states, encoder_hidden_states, temb)
+
+        # Feed-forward (GELU on concatenated text + latent)
+        ff_input = torch.cat([norm_encoder, norm_hidden], dim=1)
+        ff_output = self.ff_out(F.gelu(self.ff_proj(ff_input), approximate="tanh"))
+
+        hidden_states = hidden_states + gate_ff * ff_output[:, text_seq_length:]
+        encoder_hidden_states = encoder_hidden_states + enc_gate_ff * ff_output[:, :text_seq_length]
+
+        return hidden_states, encoder_hidden_states
+
+
+class CogVideoXTransformer3DModel(nn.Module):
+    def __init__(self,
+                 num_attention_heads=30,
+                 attention_head_dim=64,
+                 in_channels=16,
+                 out_channels=16,
+                 flip_sin_to_cos=True,
+                 freq_shift=0,
+                 time_embed_dim=512,
+                 ofs_embed_dim=None,
+                 text_embed_dim=4096,
+                 num_layers=30,
+                 dropout=0.0,
+                 attention_bias=True,
+                 sample_width=90,
+                 sample_height=60,
+                 sample_frames=49,
+                 patch_size=2,
+                 patch_size_t=None,
+                 temporal_compression_ratio=4,
+                 max_text_seq_length=226,
+                 spatial_interpolation_scale=1.875,
+                 temporal_interpolation_scale=1.0,
+                 use_rotary_positional_embeddings=False,
+                 use_learned_positional_embeddings=False,
+                 patch_bias=True,
+                 image_model=None,
+                 device=None,
+                 dtype=None,
+                 operations=None,
+                 ):
+        super().__init__()
+        self.dtype = dtype
+        dim = num_attention_heads * attention_head_dim
+        self.dim = dim
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.patch_size = patch_size
+        self.patch_size_t = patch_size_t
+        self.max_text_seq_length = max_text_seq_length
+        self.use_rotary_positional_embeddings = use_rotary_positional_embeddings
+
+        # 1. Patch embedding
+        self.patch_embed = CogVideoXPatchEmbed(
+            patch_size=patch_size,
+            patch_size_t=patch_size_t,
+            in_channels=in_channels,
+            dim=dim,
+            text_dim=text_embed_dim,
+            bias=patch_bias,
+            sample_width=sample_width,
+            sample_height=sample_height,
+            sample_frames=sample_frames,
+            temporal_compression_ratio=temporal_compression_ratio,
+            max_text_seq_length=max_text_seq_length,
+            spatial_interpolation_scale=spatial_interpolation_scale,
+            temporal_interpolation_scale=temporal_interpolation_scale,
+            use_positional_embeddings=not use_rotary_positional_embeddings,
+            use_learned_positional_embeddings=use_learned_positional_embeddings,
+            device=device, dtype=torch.float32, operations=operations,
+        )
+
+        # 2. Time embedding
+        self.time_proj_dim = dim
+        self.time_proj_flip = flip_sin_to_cos
+        self.time_proj_shift = freq_shift
+        self.time_embedding_linear_1 = operations.Linear(dim, time_embed_dim, device=device, dtype=dtype)
+        self.time_embedding_act = nn.SiLU()
+        self.time_embedding_linear_2 = operations.Linear(time_embed_dim, time_embed_dim, device=device, dtype=dtype)
+
+        # Optional OFS embedding (CogVideoX 1.5 I2V)
+        self.ofs_proj_dim = ofs_embed_dim
+        if ofs_embed_dim:
+            self.ofs_embedding_linear_1 = operations.Linear(ofs_embed_dim, ofs_embed_dim, device=device, dtype=dtype)
+            self.ofs_embedding_act = nn.SiLU()
+            self.ofs_embedding_linear_2 = operations.Linear(ofs_embed_dim, ofs_embed_dim, device=device, dtype=dtype)
+        else:
+            self.ofs_embedding_linear_1 = None
+
+        # 3. Transformer blocks
+        self.blocks = nn.ModuleList([
+            CogVideoXBlock(
+                dim=dim,
+                num_heads=num_attention_heads,
+                head_dim=attention_head_dim,
+                time_dim=time_embed_dim,
+                eps=1e-5,
+                device=device, dtype=dtype, operations=operations,
+            )
+            for _ in range(num_layers)
+        ])
+
+        self.norm_final = operations.LayerNorm(dim, eps=1e-5, elementwise_affine=True, device=device, dtype=dtype)
+
+        # 4. Output
+        self.norm_out = CogVideoXAdaLayerNorm(
+            time_dim=time_embed_dim, dim=dim, eps=1e-5,
+            device=device, dtype=dtype, operations=operations,
+        )
+
+        if patch_size_t is None:
+            output_dim = patch_size * patch_size * out_channels
+        else:
+            output_dim = patch_size * patch_size * patch_size_t * out_channels
+
+        self.proj_out = operations.Linear(dim, output_dim, device=device, dtype=dtype)
+
+        self.spatial_interpolation_scale = spatial_interpolation_scale
+        self.temporal_interpolation_scale = temporal_interpolation_scale
+        self.temporal_compression_ratio = temporal_compression_ratio
+
+    def forward(self, x, timestep, context, ofs=None, transformer_options=None, **kwargs):
+        if transformer_options is None:
+            transformer_options = {}
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, ofs, transformer_options, **kwargs)
+
+    def _forward(self, x, timestep, context, ofs=None, transformer_options=None, **kwargs):
+        if transformer_options is None:
+            transformer_options = {}
+        # ComfyUI passes [B, C, T, H, W]
+        batch_size, channels, t, h, w = x.shape
+
+        # Pad to patch size (temporal + spatial), same pattern as WAN
+        p_t = self.patch_size_t if self.patch_size_t is not None else 1
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (p_t, self.patch_size, self.patch_size))
+
+        # CogVideoX expects [B, T, C, H, W]
+        x = x.permute(0, 2, 1, 3, 4)
+        batch_size, num_frames, channels, height, width = x.shape
+
+        # Time embedding
+        t_emb = get_timestep_embedding(timestep, self.time_proj_dim, self.time_proj_flip, self.time_proj_shift)
+        t_emb = t_emb.to(dtype=x.dtype)
+        emb = self.time_embedding_linear_2(self.time_embedding_act(self.time_embedding_linear_1(t_emb)))
+
+        if self.ofs_embedding_linear_1 is not None and ofs is not None:
+            ofs_emb = get_timestep_embedding(ofs, self.ofs_proj_dim, self.time_proj_flip, self.time_proj_shift)
+            ofs_emb = ofs_emb.to(dtype=x.dtype)
+            ofs_emb = self.ofs_embedding_linear_2(self.ofs_embedding_act(self.ofs_embedding_linear_1(ofs_emb)))
+            emb = emb + ofs_emb
+
+        # Patch embedding
+        hidden_states = self.patch_embed(context, x)
+
+        text_seq_length = context.shape[1]
+        encoder_hidden_states = hidden_states[:, :text_seq_length]
+        hidden_states = hidden_states[:, text_seq_length:]
+
+        # Rotary embeddings (if used)
+        image_rotary_emb = None
+        if self.use_rotary_positional_embeddings:
+            post_patch_height = height // self.patch_size
+            post_patch_width = width // self.patch_size
+            if self.patch_size_t is None:
+                post_time = num_frames
+            else:
+                post_time = num_frames // self.patch_size_t
+            image_rotary_emb = self._get_rotary_emb(post_patch_height, post_patch_width, post_time, device=x.device)
+
+        # Transformer blocks
+        for i, block in enumerate(self.blocks):
+            hidden_states, encoder_hidden_states = block(
+                hidden_states=hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                temb=emb,
+                image_rotary_emb=image_rotary_emb,
+                transformer_options=transformer_options,
+            )
+
+        hidden_states = self.norm_final(hidden_states)
+
+        # Output projection
+        hidden_states = self.norm_out(hidden_states, temb=emb)
+        hidden_states = self.proj_out(hidden_states)
+
+        # Unpatchify
+        p = self.patch_size
+        p_t = self.patch_size_t
+
+        if p_t is None:
+            output = hidden_states.reshape(batch_size, num_frames, height // p, width // p, -1, p, p)
+            output = output.permute(0, 1, 4, 2, 5, 3, 6).flatten(5, 6).flatten(3, 4)
+        else:
+            output = hidden_states.reshape(
+                batch_size, (num_frames + p_t - 1) // p_t, height // p, width // p, -1, p_t, p, p
+            )
+            output = output.permute(0, 1, 5, 4, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(1, 2)
+
+        # Back to ComfyUI format [B, C, T, H, W] and crop padding
+        output = output.permute(0, 2, 1, 3, 4)[:, :, :t, :h, :w]
+        return output
+
+    def _get_rotary_emb(self, h, w, t, device):
+        """Compute CogVideoX 3D rotary positional embeddings.
+
+        For CogVideoX 1.5 (patch_size_t != None): uses "slice" mode — grid positions
+        are integer arange computed at max_size, then sliced to actual size.
+        For CogVideoX 1.0 (patch_size_t == None): uses "linspace" mode with crop coords
+        scaled by spatial_interpolation_scale.
+        """
+        d = self.attention_head_dim
+        dim_t = d // 4
+        dim_h = d // 8 * 3
+        dim_w = d // 8 * 3
+
+        if self.patch_size_t is not None:
+            # CogVideoX 1.5: "slice" mode — positions are simple integer indices
+            # Compute at max(sample_size, actual_size) then slice to actual
+            base_h = self.patch_embed.sample_height // self.patch_size
+            base_w = self.patch_embed.sample_width // self.patch_size
+            max_h = max(base_h, h)
+            max_w = max(base_w, w)
+
+            grid_h = torch.arange(max_h, device=device, dtype=torch.float32)
+            grid_w = torch.arange(max_w, device=device, dtype=torch.float32)
+            grid_t = torch.arange(t, device=device, dtype=torch.float32)
+        else:
+            # CogVideoX 1.0: "linspace" mode with interpolation scale
+            grid_h = torch.linspace(0, h - 1, h, device=device, dtype=torch.float32) * self.spatial_interpolation_scale
+            grid_w = torch.linspace(0, w - 1, w, device=device, dtype=torch.float32) * self.spatial_interpolation_scale
+            grid_t = torch.arange(t, device=device, dtype=torch.float32)
+
+        freqs_t = _get_1d_rotary_pos_embed(dim_t, grid_t)
+        freqs_h = _get_1d_rotary_pos_embed(dim_h, grid_h)
+        freqs_w = _get_1d_rotary_pos_embed(dim_w, grid_w)
+
+        t_cos, t_sin = freqs_t
+        h_cos, h_sin = freqs_h
+        w_cos, w_sin = freqs_w
+
+        # Slice to actual size (for "slice" mode where grids may be larger)
+        t_cos, t_sin = t_cos[:t], t_sin[:t]
+        h_cos, h_sin = h_cos[:h], h_sin[:h]
+        w_cos, w_sin = w_cos[:w], w_sin[:w]
+
+        # Broadcast and concatenate into [T*H*W, head_dim]
+        t_cos = t_cos[:, None, None, :].expand(-1, h, w, -1)
+        t_sin = t_sin[:, None, None, :].expand(-1, h, w, -1)
+        h_cos = h_cos[None, :, None, :].expand(t, -1, w, -1)
+        h_sin = h_sin[None, :, None, :].expand(t, -1, w, -1)
+        w_cos = w_cos[None, None, :, :].expand(t, h, -1, -1)
+        w_sin = w_sin[None, None, :, :].expand(t, h, -1, -1)
+
+        cos = torch.cat([t_cos, h_cos, w_cos], dim=-1).reshape(t * h * w, -1)
+        sin = torch.cat([t_sin, h_sin, w_sin], dim=-1).reshape(t * h * w, -1)
+        return (cos, sin)

comfy/ldm/cogvideo/vae.py

@@ -0,0 +1,566 @@
+# CogVideoX VAE - ported to ComfyUI native ops
+# Architecture reference: diffusers AutoencoderKLCogVideoX
+# Style reference: comfy/ldm/wan/vae.py
+
+import numpy as np
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+
+class CausalConv3d(nn.Module):
+    """Causal 3D convolution with temporal padding.
+
+    Uses comfy.ops.Conv3d with autopad='causal_zero' fast path: when input has
+    a single temporal frame and no cache, the 3D conv weight is sliced to act
+    as a 2D conv, avoiding computation on zero-padded temporal dimensions.
+    """
+    def __init__(self, in_channels, out_channels, kernel_size, stride=1, dilation=1, pad_mode="constant"):
+        super().__init__()
+        if isinstance(kernel_size, int):
+            kernel_size = (kernel_size,) * 3
+
+        time_kernel, height_kernel, width_kernel = kernel_size
+        self.time_kernel_size = time_kernel
+        self.pad_mode = pad_mode
+
+        height_pad = (height_kernel - 1) // 2
+        width_pad = (width_kernel - 1) // 2
+        self.time_causal_padding = (width_pad, width_pad, height_pad, height_pad, time_kernel - 1, 0)
+
+        stride = stride if isinstance(stride, tuple) else (stride, 1, 1)
+        dilation = (dilation, 1, 1)
+        self.conv = ops.Conv3d(
+            in_channels, out_channels, kernel_size,
+            stride=stride, dilation=dilation,
+            padding=(0, height_pad, width_pad),
+        )
+
+    def forward(self, x, conv_cache=None):
+        if self.pad_mode == "replicate":
+            x = F.pad(x, self.time_causal_padding, mode="replicate")
+            conv_cache = None
+        else:
+            kernel_t = self.time_kernel_size
+            if kernel_t > 1:
+                if conv_cache is None and x.shape[2] == 1:
+                    # Fast path: single frame, no cache. All temporal padding
+                    # frames are copies of the input (replicate-style), so the
+                    # 3D conv reduces to a 2D conv with summed temporal kernel.
+                    w = comfy.ops.cast_to_input(self.conv.weight, x)
+                    b = comfy.ops.cast_to_input(self.conv.bias, x) if self.conv.bias is not None else None
+                    w2d = w.sum(dim=2, keepdim=True)
+                    out = F.conv3d(x, w2d, b,
+                                   self.conv.stride, self.conv.padding,
+                                   self.conv.dilation, self.conv.groups)
+                    return out, None
+                cached = [conv_cache] if conv_cache is not None else [x[:, :, :1]] * (kernel_t - 1)
+                x = torch.cat(cached + [x], dim=2)
+            conv_cache = x[:, :, -self.time_kernel_size + 1:].clone() if self.time_kernel_size > 1 else None
+
+        out = self.conv(x)
+        return out, conv_cache
+
+
+def _interpolate_zq(zq, target_size):
+    """Interpolate latent z to target (T, H, W), matching CogVideoX's first-frame-special handling."""
+    t = target_size[0]
+    if t > 1 and t % 2 == 1:
+        z_first = F.interpolate(zq[:, :, :1], size=(1, target_size[1], target_size[2]))
+        z_rest = F.interpolate(zq[:, :, 1:], size=(t - 1, target_size[1], target_size[2]))
+        return torch.cat([z_first, z_rest], dim=2)
+    return F.interpolate(zq, size=target_size)
+
+
+class SpatialNorm3D(nn.Module):
+    """Spatially conditioned normalization."""
+    def __init__(self, f_channels, zq_channels, groups=32):
+        super().__init__()
+        self.norm_layer = ops.GroupNorm(num_channels=f_channels, num_groups=groups, eps=1e-6, affine=True)
+        self.conv_y = CausalConv3d(zq_channels, f_channels, kernel_size=1, stride=1)
+        self.conv_b = CausalConv3d(zq_channels, f_channels, kernel_size=1, stride=1)
+
+    def forward(self, f, zq, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+
+        if zq.shape[-3:] != f.shape[-3:]:
+            zq = _interpolate_zq(zq, f.shape[-3:])
+
+        conv_y, new_cache["conv_y"] = self.conv_y(zq, conv_cache=conv_cache.get("conv_y"))
+        conv_b, new_cache["conv_b"] = self.conv_b(zq, conv_cache=conv_cache.get("conv_b"))
+
+        return self.norm_layer(f) * conv_y + conv_b, new_cache
+
+
+class ResnetBlock3D(nn.Module):
+    """3D ResNet block with optional spatial norm."""
+    def __init__(self, in_channels, out_channels=None, temb_channels=512, groups=32,
+                 eps=1e-6, act_fn="silu", spatial_norm_dim=None, pad_mode="first"):
+        super().__init__()
+        out_channels = out_channels or in_channels
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.spatial_norm_dim = spatial_norm_dim
+
+        if act_fn == "silu":
+            self.nonlinearity = nn.SiLU()
+        elif act_fn == "swish":
+            self.nonlinearity = nn.SiLU()
+        else:
+            self.nonlinearity = nn.SiLU()
+
+        if spatial_norm_dim is None:
+            self.norm1 = ops.GroupNorm(num_channels=in_channels, num_groups=groups, eps=eps)
+            self.norm2 = ops.GroupNorm(num_channels=out_channels, num_groups=groups, eps=eps)
+        else:
+            self.norm1 = SpatialNorm3D(in_channels, spatial_norm_dim, groups=groups)
+            self.norm2 = SpatialNorm3D(out_channels, spatial_norm_dim, groups=groups)
+
+        self.conv1 = CausalConv3d(in_channels, out_channels, kernel_size=3, pad_mode=pad_mode)
+
+        if temb_channels > 0:
+            self.temb_proj = ops.Linear(temb_channels, out_channels)
+
+        self.conv2 = CausalConv3d(out_channels, out_channels, kernel_size=3, pad_mode=pad_mode)
+
+        if in_channels != out_channels:
+            self.conv_shortcut = ops.Conv3d(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
+        else:
+            self.conv_shortcut = None
+
+    def forward(self, x, temb=None, zq=None, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+        residual = x
+
+        if zq is not None:
+            x, new_cache["norm1"] = self.norm1(x, zq, conv_cache=conv_cache.get("norm1"))
+        else:
+            x = self.norm1(x)
+
+        x = self.nonlinearity(x)
+        x, new_cache["conv1"] = self.conv1(x, conv_cache=conv_cache.get("conv1"))
+
+        if temb is not None and hasattr(self, "temb_proj"):
+            x = x + self.temb_proj(self.nonlinearity(temb))[:, :, None, None, None]
+
+        if zq is not None:
+            x, new_cache["norm2"] = self.norm2(x, zq, conv_cache=conv_cache.get("norm2"))
+        else:
+            x = self.norm2(x)
+
+        x = self.nonlinearity(x)
+        x, new_cache["conv2"] = self.conv2(x, conv_cache=conv_cache.get("conv2"))
+
+        if self.conv_shortcut is not None:
+            residual = self.conv_shortcut(residual)
+
+        return x + residual, new_cache
+
+
+class Downsample3D(nn.Module):
+    """3D downsampling with optional temporal compression."""
+    def __init__(self, in_channels, out_channels, kernel_size=3, stride=2, padding=0, compress_time=False):
+        super().__init__()
+        self.conv = ops.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding)
+        self.compress_time = compress_time
+
+    def forward(self, x):
+        if self.compress_time:
+            b, c, t, h, w = x.shape
+            x = x.permute(0, 3, 4, 1, 2).reshape(b * h * w, c, t)
+            if t % 2 == 1:
+                x_first, x_rest = x[..., 0], x[..., 1:]
+                if x_rest.shape[-1] > 0:
+                    x_rest = F.avg_pool1d(x_rest, kernel_size=2, stride=2)
+                x = torch.cat([x_first[..., None], x_rest], dim=-1)
+                x = x.reshape(b, h, w, c, x.shape[-1]).permute(0, 3, 4, 1, 2)
+            else:
+                x = F.avg_pool1d(x, kernel_size=2, stride=2)
+                x = x.reshape(b, h, w, c, x.shape[-1]).permute(0, 3, 4, 1, 2)
+
+        pad = (0, 1, 0, 1)
+        x = F.pad(x, pad, mode="constant", value=0)
+        b, c, t, h, w = x.shape
+        x = x.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
+        x = self.conv(x)
+        x = x.reshape(b, t, x.shape[1], x.shape[2], x.shape[3]).permute(0, 2, 1, 3, 4)
+        return x
+
+
+class Upsample3D(nn.Module):
+    """3D upsampling with optional temporal decompression."""
+    def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, padding=1, compress_time=False):
+        super().__init__()
+        self.conv = ops.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding)
+        self.compress_time = compress_time
+
+    def forward(self, x):
+        if self.compress_time:
+            if x.shape[2] > 1 and x.shape[2] % 2 == 1:
+                x_first, x_rest = x[:, :, 0], x[:, :, 1:]
+                x_first = F.interpolate(x_first, scale_factor=2.0)
+                x_rest = F.interpolate(x_rest, scale_factor=2.0)
+                x = torch.cat([x_first[:, :, None, :, :], x_rest], dim=2)
+            elif x.shape[2] > 1:
+                x = F.interpolate(x, scale_factor=2.0)
+            else:
+                x = x.squeeze(2)
+                x = F.interpolate(x, scale_factor=2.0)
+                x = x[:, :, None, :, :]
+        else:
+            b, c, t, h, w = x.shape
+            x = x.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
+            x = F.interpolate(x, scale_factor=2.0)
+            x = x.reshape(b, t, c, *x.shape[2:]).permute(0, 2, 1, 3, 4)
+
+        b, c, t, h, w = x.shape
+        x = x.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
+        x = self.conv(x)
+        x = x.reshape(b, t, *x.shape[1:]).permute(0, 2, 1, 3, 4)
+        return x
+
+
+class DownBlock3D(nn.Module):
+    def __init__(self, in_channels, out_channels, temb_channels=0, num_layers=1,
+                 eps=1e-6, act_fn="silu", groups=32, add_downsample=True,
+                 compress_time=False, pad_mode="first"):
+        super().__init__()
+        self.resnets = nn.ModuleList([
+            ResnetBlock3D(
+                in_channels=in_channels if i == 0 else out_channels,
+                out_channels=out_channels,
+                temb_channels=temb_channels,
+                groups=groups, eps=eps, act_fn=act_fn, pad_mode=pad_mode,
+            )
+            for i in range(num_layers)
+        ])
+        self.downsamplers = nn.ModuleList([Downsample3D(out_channels, out_channels, compress_time=compress_time)]) if add_downsample else None
+
+    def forward(self, x, temb=None, zq=None, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+        for i, resnet in enumerate(self.resnets):
+            x, new_cache[f"resnet_{i}"] = resnet(x, temb, zq, conv_cache=conv_cache.get(f"resnet_{i}"))
+        if self.downsamplers is not None:
+            for ds in self.downsamplers:
+                x = ds(x)
+        return x, new_cache
+
+
+class MidBlock3D(nn.Module):
+    def __init__(self, in_channels, temb_channels=0, num_layers=1,
+                 eps=1e-6, act_fn="silu", groups=32, spatial_norm_dim=None, pad_mode="first"):
+        super().__init__()
+        self.resnets = nn.ModuleList([
+            ResnetBlock3D(
+                in_channels=in_channels, out_channels=in_channels,
+                temb_channels=temb_channels, groups=groups, eps=eps,
+                act_fn=act_fn, spatial_norm_dim=spatial_norm_dim, pad_mode=pad_mode,
+            )
+            for _ in range(num_layers)
+        ])
+
+    def forward(self, x, temb=None, zq=None, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+        for i, resnet in enumerate(self.resnets):
+            x, new_cache[f"resnet_{i}"] = resnet(x, temb, zq, conv_cache=conv_cache.get(f"resnet_{i}"))
+        return x, new_cache
+
+
+class UpBlock3D(nn.Module):
+    def __init__(self, in_channels, out_channels, temb_channels=0, num_layers=1,
+                 eps=1e-6, act_fn="silu", groups=32, spatial_norm_dim=16,
+                 add_upsample=True, compress_time=False, pad_mode="first"):
+        super().__init__()
+        self.resnets = nn.ModuleList([
+            ResnetBlock3D(
+                in_channels=in_channels if i == 0 else out_channels,
+                out_channels=out_channels,
+                temb_channels=temb_channels, groups=groups, eps=eps,
+                act_fn=act_fn, spatial_norm_dim=spatial_norm_dim, pad_mode=pad_mode,
+            )
+            for i in range(num_layers)
+        ])
+        self.upsamplers = nn.ModuleList([Upsample3D(out_channels, out_channels, compress_time=compress_time)]) if add_upsample else None
+
+    def forward(self, x, temb=None, zq=None, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+        for i, resnet in enumerate(self.resnets):
+            x, new_cache[f"resnet_{i}"] = resnet(x, temb, zq, conv_cache=conv_cache.get(f"resnet_{i}"))
+        if self.upsamplers is not None:
+            for us in self.upsamplers:
+                x = us(x)
+        return x, new_cache
+
+
+class Encoder3D(nn.Module):
+    def __init__(self, in_channels=3, out_channels=16,
+                 block_out_channels=(128, 256, 256, 512),
+                 layers_per_block=3, act_fn="silu",
+                 eps=1e-6, groups=32, pad_mode="first",
+                 temporal_compression_ratio=4):
+        super().__init__()
+        temporal_compress_level = int(np.log2(temporal_compression_ratio))
+
+        self.conv_in = CausalConv3d(in_channels, block_out_channels[0], kernel_size=3, pad_mode=pad_mode)
+
+        self.down_blocks = nn.ModuleList()
+        output_channel = block_out_channels[0]
+        for i in range(len(block_out_channels)):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final = i == len(block_out_channels) - 1
+            compress_time = i < temporal_compress_level
+
+            self.down_blocks.append(DownBlock3D(
+                in_channels=input_channel, out_channels=output_channel,
+                temb_channels=0, num_layers=layers_per_block,
+                eps=eps, act_fn=act_fn, groups=groups,
+                add_downsample=not is_final, compress_time=compress_time,
+            ))
+
+        self.mid_block = MidBlock3D(
+            in_channels=block_out_channels[-1], temb_channels=0,
+            num_layers=2, eps=eps, act_fn=act_fn, groups=groups, pad_mode=pad_mode,
+        )
+
+        self.norm_out = ops.GroupNorm(groups, block_out_channels[-1], eps=1e-6)
+        self.conv_act = nn.SiLU()
+        self.conv_out = CausalConv3d(block_out_channels[-1], 2 * out_channels, kernel_size=3, pad_mode=pad_mode)
+
+    def forward(self, x, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+
+        x, new_cache["conv_in"] = self.conv_in(x, conv_cache=conv_cache.get("conv_in"))
+
+        for i, block in enumerate(self.down_blocks):
+            key = f"down_block_{i}"
+            x, new_cache[key] = block(x, None, None, conv_cache.get(key))
+
+        x, new_cache["mid_block"] = self.mid_block(x, None, None, conv_cache=conv_cache.get("mid_block"))
+
+        x = self.norm_out(x)
+        x = self.conv_act(x)
+        x, new_cache["conv_out"] = self.conv_out(x, conv_cache=conv_cache.get("conv_out"))
+
+        return x, new_cache
+
+
+class Decoder3D(nn.Module):
+    def __init__(self, in_channels=16, out_channels=3,
+                 block_out_channels=(128, 256, 256, 512),
+                 layers_per_block=3, act_fn="silu",
+                 eps=1e-6, groups=32, pad_mode="first",
+                 temporal_compression_ratio=4):
+        super().__init__()
+        reversed_channels = list(reversed(block_out_channels))
+        temporal_compress_level = int(np.log2(temporal_compression_ratio))
+
+        self.conv_in = CausalConv3d(in_channels, reversed_channels[0], kernel_size=3, pad_mode=pad_mode)
+
+        self.mid_block = MidBlock3D(
+            in_channels=reversed_channels[0], temb_channels=0,
+            num_layers=2, eps=eps, act_fn=act_fn, groups=groups,
+            spatial_norm_dim=in_channels, pad_mode=pad_mode,
+        )
+
+        self.up_blocks = nn.ModuleList()
+        output_channel = reversed_channels[0]
+        for i in range(len(block_out_channels)):
+            prev_channel = output_channel
+            output_channel = reversed_channels[i]
+            is_final = i == len(block_out_channels) - 1
+            compress_time = i < temporal_compress_level
+
+            self.up_blocks.append(UpBlock3D(
+                in_channels=prev_channel, out_channels=output_channel,
+                temb_channels=0, num_layers=layers_per_block + 1,
+                eps=eps, act_fn=act_fn, groups=groups,
+                spatial_norm_dim=in_channels,
+                add_upsample=not is_final, compress_time=compress_time,
+            ))
+
+        self.norm_out = SpatialNorm3D(reversed_channels[-1], in_channels, groups=groups)
+        self.conv_act = nn.SiLU()
+        self.conv_out = CausalConv3d(reversed_channels[-1], out_channels, kernel_size=3, pad_mode=pad_mode)
+
+    def forward(self, sample, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+
+        x, new_cache["conv_in"] = self.conv_in(sample, conv_cache=conv_cache.get("conv_in"))
+
+        x, new_cache["mid_block"] = self.mid_block(x, None, sample, conv_cache=conv_cache.get("mid_block"))
+
+        for i, block in enumerate(self.up_blocks):
+            key = f"up_block_{i}"
+            x, new_cache[key] = block(x, None, sample, conv_cache=conv_cache.get(key))
+
+        x, new_cache["norm_out"] = self.norm_out(x, sample, conv_cache=conv_cache.get("norm_out"))
+        x = self.conv_act(x)
+        x, new_cache["conv_out"] = self.conv_out(x, conv_cache=conv_cache.get("conv_out"))
+
+        return x, new_cache
+
+
+
+class AutoencoderKLCogVideoX(nn.Module):
+    """CogVideoX VAE. Spatial tiling/slicing handled by ComfyUI's VAE wrapper.
+
+    Uses rolling temporal decode: conv_in + mid_block + temporal up_blocks run
+    on the full (low-res) tensor, then the expensive spatial-only up_blocks +
+    norm_out + conv_out are processed in small temporal chunks with conv_cache
+    carrying causal state between chunks. This keeps peak VRAM proportional to
+    chunk_size rather than total frame count.
+    """
+
+    def __init__(self,
+                 in_channels=3, out_channels=3,
+                 block_out_channels=(128, 256, 256, 512),
+                 latent_channels=16, layers_per_block=3,
+                 act_fn="silu", eps=1e-6, groups=32,
+                 temporal_compression_ratio=4,
+                 ):
+        super().__init__()
+        self.latent_channels = latent_channels
+        self.temporal_compression_ratio = temporal_compression_ratio
+
+        self.encoder = Encoder3D(
+            in_channels=in_channels, out_channels=latent_channels,
+            block_out_channels=block_out_channels, layers_per_block=layers_per_block,
+            act_fn=act_fn, eps=eps, groups=groups,
+            temporal_compression_ratio=temporal_compression_ratio,
+        )
+        self.decoder = Decoder3D(
+            in_channels=latent_channels, out_channels=out_channels,
+            block_out_channels=block_out_channels, layers_per_block=layers_per_block,
+            act_fn=act_fn, eps=eps, groups=groups,
+            temporal_compression_ratio=temporal_compression_ratio,
+        )
+
+        self.num_latent_frames_batch_size = 2
+        self.num_sample_frames_batch_size = 8
+
+    def encode(self, x):
+        t = x.shape[2]
+        frame_batch = self.num_sample_frames_batch_size
+        remainder = t % frame_batch
+        conv_cache = None
+        enc = []
+
+        # Process remainder frames first so only the first chunk can have an
+        # odd temporal dimension — where Downsample3D's first-frame-special
+        # handling in temporal compression is actually correct.
+        if remainder > 0:
+            chunk, conv_cache = self.encoder(x[:, :, :remainder], conv_cache=conv_cache)
+            enc.append(chunk.to(x.device))
+
+        for start in range(remainder, t, frame_batch):
+            chunk, conv_cache = self.encoder(x[:, :, start:start + frame_batch], conv_cache=conv_cache)
+            enc.append(chunk.to(x.device))
+
+        enc = torch.cat(enc, dim=2)
+        mean, _ = enc.chunk(2, dim=1)
+        return mean
+
+    def decode(self, z):
+        return self._decode_rolling(z)
+
+    def _decode_batched(self, z):
+        """Original batched decode - processes 2 latent frames through full decoder."""
+        t = z.shape[2]
+        frame_batch = self.num_latent_frames_batch_size
+        num_batches = max(t // frame_batch, 1)
+        conv_cache = None
+        dec = []
+        for i in range(num_batches):
+            remaining = t % frame_batch
+            start = frame_batch * i + (0 if i == 0 else remaining)
+            end = frame_batch * (i + 1) + remaining
+            chunk, conv_cache = self.decoder(z[:, :, start:end], conv_cache=conv_cache)
+            dec.append(chunk.cpu())
+        return torch.cat(dec, dim=2).to(z.device)
+
+    def _decode_rolling(self, z):
+        """Rolling decode - processes low-res layers on full tensor, then rolls
+        through expensive high-res layers in temporal chunks."""
+        decoder = self.decoder
+        device = z.device
+
+        # Determine which up_blocks have temporal upsample vs spatial-only.
+        # Temporal up_blocks are cheap (low res), spatial-only are expensive.
+        temporal_compress_level = int(np.log2(self.temporal_compression_ratio))
+        split_at = temporal_compress_level  # first N up_blocks do temporal upsample
+
+        # Phase 1: conv_in + mid_block + temporal up_blocks on full tensor (low/medium res)
+        x, _ = decoder.conv_in(z)
+        x, _ = decoder.mid_block(x, None, z)
+
+        for i in range(split_at):
+            x, _ = decoder.up_blocks[i](x, None, z)
+
+        # Phase 2: remaining spatial-only up_blocks + norm_out + conv_out in temporal chunks
+        remaining_blocks = list(range(split_at, len(decoder.up_blocks)))
+        chunk_size = 4  # pixel frames per chunk through high-res layers
+        t_expanded = x.shape[2]
+
+        if t_expanded <= chunk_size or len(remaining_blocks) == 0:
+            # Small enough to process in one go
+            for i in remaining_blocks:
+                x, _ = decoder.up_blocks[i](x, None, z)
+            x, _ = decoder.norm_out(x, z)
+            x = decoder.conv_act(x)
+            x, _ = decoder.conv_out(x)
+            return x
+
+        # Expand z temporally once to match Phase 2's time dimension.
+        # z stays at latent spatial resolution so this is small (~16 MB vs ~1.3 GB
+        # for the old approach of pre-interpolating to every pixel resolution).
+        z_time_expanded = _interpolate_zq(z, (t_expanded, z.shape[3], z.shape[4]))
+
+        # Process in temporal chunks, interpolating spatially per-chunk to avoid
+        # allocating full [B, C, t_expanded, H, W] tensors at each resolution.
+        dec_out = []
+        conv_caches = {}
+
+        for chunk_start in range(0, t_expanded, chunk_size):
+            chunk_end = min(chunk_start + chunk_size, t_expanded)
+            x_chunk = x[:, :, chunk_start:chunk_end]
+            z_t_chunk = z_time_expanded[:, :, chunk_start:chunk_end]
+            z_spatial_cache = {}
+
+            for i in remaining_blocks:
+                block = decoder.up_blocks[i]
+                cache_key = f"up_block_{i}"
+                hw_key = (x_chunk.shape[3], x_chunk.shape[4])
+                if hw_key not in z_spatial_cache:
+                    if z_t_chunk.shape[3] == hw_key[0] and z_t_chunk.shape[4] == hw_key[1]:
+                        z_spatial_cache[hw_key] = z_t_chunk
+                    else:
+                        z_spatial_cache[hw_key] = F.interpolate(z_t_chunk, size=(z_t_chunk.shape[2], hw_key[0], hw_key[1]))
+                x_chunk, new_cache = block(x_chunk, None, z_spatial_cache[hw_key], conv_cache=conv_caches.get(cache_key))
+                conv_caches[cache_key] = new_cache
+
+            hw_key = (x_chunk.shape[3], x_chunk.shape[4])
+            if hw_key not in z_spatial_cache:
+                z_spatial_cache[hw_key] = F.interpolate(z_t_chunk, size=(z_t_chunk.shape[2], hw_key[0], hw_key[1]))
+            x_chunk, new_cache = decoder.norm_out(x_chunk, z_spatial_cache[hw_key], conv_cache=conv_caches.get("norm_out"))
+            conv_caches["norm_out"] = new_cache
+            x_chunk = decoder.conv_act(x_chunk)
+            x_chunk, new_cache = decoder.conv_out(x_chunk, conv_cache=conv_caches.get("conv_out"))
+            conv_caches["conv_out"] = new_cache
+
+            dec_out.append(x_chunk.cpu())
+            del z_spatial_cache
+
+        del x, z_time_expanded
+        return torch.cat(dec_out, dim=2).to(device)

comfy/ldm/lightricks/av_model.py

@@ -16,6 +16,7 @@ from comfy.ldm.lightricks.model import (
 from comfy.ldm.lightricks.symmetric_patchifier import AudioPatchifier
 from comfy.ldm.lightricks.embeddings_connector import Embeddings1DConnector
 import comfy.ldm.common_dit
+import comfy.model_prefetch
 
 class CompressedTimestep:
     """Store video timestep embeddings in compressed form using per-frame indexing."""
@@ -907,9 +908,11 @@ class LTXAVModel(LTXVModel):
         """Process transformer blocks for LTXAV."""
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
+        prefetch_queue = comfy.model_prefetch.make_prefetch_queue(list(self.transformer_blocks), vx.device, transformer_options)
 
         # Process transformer blocks
         for i, block in enumerate(self.transformer_blocks):
+            comfy.model_prefetch.prefetch_queue_pop(prefetch_queue, vx.device, block)
             if ("double_block", i) in blocks_replace:
 
                 def block_wrap(args):
@@ -982,6 +985,8 @@ class LTXAVModel(LTXVModel):
                     a_prompt_timestep=a_prompt_timestep,
                 )
 
+        comfy.model_prefetch.prefetch_queue_pop(prefetch_queue, vx.device, None)
+
         return [vx, ax]
 
     def _process_output(self, x, embedded_timestep, keyframe_idxs, **kwargs):

comfy/ldm/modules/attention.py

@@ -14,6 +14,8 @@ from .sub_quadratic_attention import efficient_dot_product_attention
 
 from comfy import model_management
 
+TORCH_HAS_GQA = model_management.torch_version_numeric >= (2, 5)
+
 if model_management.xformers_enabled():
     import xformers
     import xformers.ops
@@ -150,7 +152,12 @@ def attention_basic(q, k, v, heads, mask=None, attn_precision=None, skip_reshape
         b, _, dim_head = q.shape
         dim_head //= heads
 
-    scale = dim_head ** -0.5
+    if kwargs.get("enable_gqa", False) and q.shape[-3] != k.shape[-3]:
+        n_rep = q.shape[-3] // k.shape[-3]
+        k = k.repeat_interleave(n_rep, dim=-3)
+        v = v.repeat_interleave(n_rep, dim=-3)
+
+    scale = kwargs.get("scale", dim_head ** -0.5)
 
     h = heads
     if skip_reshape:
@@ -219,6 +226,10 @@ def attention_sub_quad(query, key, value, heads, mask=None, attn_precision=None,
         b, _, dim_head = query.shape
         dim_head //= heads
 
+    if "scale" in kwargs:
+        # Pre-scale query to match requested scale (cancels internal 1/sqrt(dim_head))
+        query = query * (kwargs["scale"] * dim_head ** 0.5)
+
     if skip_reshape:
         query = query.reshape(b * heads, -1, dim_head)
         value = value.reshape(b * heads, -1, dim_head)
@@ -290,7 +301,7 @@ def attention_split(q, k, v, heads, mask=None, attn_precision=None, skip_reshape
         b, _, dim_head = q.shape
         dim_head //= heads
 
-    scale = dim_head ** -0.5
+    scale = kwargs.get("scale", dim_head ** -0.5)
 
     if skip_reshape:
          q, k, v = map(
@@ -500,8 +511,13 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
         if mask.ndim == 3:
             mask = mask.unsqueeze(1)
 
+    # Pass through extra SDPA kwargs (scale, enable_gqa) if provided
+    # enable_gqa requires PyTorch 2.5+; older versions use manual KV expansion above
+    sdpa_keys = ("scale", "enable_gqa") if TORCH_HAS_GQA else ("scale",)
+    sdpa_extra = {k: v for k, v in kwargs.items() if k in sdpa_keys}
+
     if SDP_BATCH_LIMIT >= b:
-        out = comfy.ops.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
+        out = comfy.ops.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False, **sdpa_extra)
         if not skip_output_reshape:
             out = (
                 out.transpose(1, 2).reshape(b, -1, heads * dim_head)
@@ -519,7 +535,7 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
                 k[i : i + SDP_BATCH_LIMIT],
                 v[i : i + SDP_BATCH_LIMIT],
                 attn_mask=m,
-                dropout_p=0.0, is_causal=False
+                dropout_p=0.0, is_causal=False, **sdpa_extra
             ).transpose(1, 2).reshape(-1, q.shape[2], heads * dim_head)
     return out
 

comfy/ldm/sam3/detector.py

@@ -0,0 +1,596 @@
+# SAM3 detector: transformer encoder-decoder, segmentation head, geometry encoder, scoring.
+
+import math
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torchvision.ops import roi_align
+
+from comfy.ldm.modules.attention import optimized_attention
+from comfy.ldm.sam3.tracker import SAM3Tracker, SAM31Tracker
+from comfy.ldm.sam3.sam import SAM3VisionBackbone  # noqa: used in __init__
+from comfy.ldm.sam3.sam import MLP, PositionEmbeddingSine
+
+TRACKER_CLASSES = {"SAM3": SAM3Tracker, "SAM31": SAM31Tracker}
+from comfy.ops import cast_to_input
+
+
+def box_cxcywh_to_xyxy(x):
+    cx, cy, w, h = x.unbind(-1)
+    return torch.stack([cx - 0.5 * w, cy - 0.5 * h, cx + 0.5 * w, cy + 0.5 * h], dim=-1)
+
+
+def gen_sineembed_for_position(pos_tensor, num_feats=256):
+    """Per-coordinate sinusoidal embedding: (..., N) -> (..., N * num_feats)."""
+    assert num_feats % 2 == 0
+    hdim = num_feats // 2
+    freqs = 10000.0 ** (2 * (torch.arange(hdim, dtype=torch.float32, device=pos_tensor.device) // 2) / hdim)
+    embeds = []
+    for c in range(pos_tensor.shape[-1]):
+        raw = (pos_tensor[..., c].float() * 2 * math.pi).unsqueeze(-1) / freqs
+        embeds.append(torch.stack([raw[..., 0::2].sin(), raw[..., 1::2].cos()], dim=-1).flatten(-2))
+    return torch.cat(embeds, dim=-1).to(pos_tensor.dtype)
+
+
+class SplitMHA(nn.Module):
+    """Multi-head attention with separate Q/K/V projections (split from fused in_proj_weight)."""
+    def __init__(self, d_model, num_heads=8, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.num_heads = num_heads
+        self.q_proj = operations.Linear(d_model, d_model, device=device, dtype=dtype)
+        self.k_proj = operations.Linear(d_model, d_model, device=device, dtype=dtype)
+        self.v_proj = operations.Linear(d_model, d_model, device=device, dtype=dtype)
+        self.out_proj = operations.Linear(d_model, d_model, device=device, dtype=dtype)
+
+    def forward(self, q_input, k_input=None, v_input=None, mask=None):
+        q = self.q_proj(q_input)
+        if k_input is None:
+            k = self.k_proj(q_input)
+            v = self.v_proj(q_input)
+        else:
+            k = self.k_proj(k_input)
+            v = self.v_proj(v_input if v_input is not None else k_input)
+        if mask is not None and mask.ndim == 2:
+            mask = mask[:, None, None, :]  # [B, T] -> [B, 1, 1, T] for SDPA broadcast
+        dtype = q.dtype  # manual_cast may produce mixed dtypes
+        out = optimized_attention(q, k.to(dtype), v.to(dtype), self.num_heads, mask=mask, low_precision_attention=False)
+        return self.out_proj(out)
+
+
+class MLPWithNorm(nn.Module):
+    """MLP with residual connection and output LayerNorm."""
+    def __init__(self, input_dim, hidden_dim, output_dim, num_layers, residual=True, device=None, dtype=None, operations=None):
+        super().__init__()
+        dims = [input_dim] + [hidden_dim] * (num_layers - 1) + [output_dim]
+        self.layers = nn.ModuleList([
+            operations.Linear(dims[i], dims[i + 1], device=device, dtype=dtype)
+            for i in range(num_layers)
+        ])
+        self.out_norm = operations.LayerNorm(output_dim, device=device, dtype=dtype)
+        self.residual = residual and (input_dim == output_dim)
+
+    def forward(self, x):
+        orig = x
+        for i, layer in enumerate(self.layers):
+            x = layer(x)
+            if i < len(self.layers) - 1:
+                x = F.relu(x)
+        if self.residual:
+            x = x + orig
+        return self.out_norm(x)
+
+
+class EncoderLayer(nn.Module):
+    def __init__(self, d_model=256, num_heads=8, dim_ff=2048, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.self_attn = SplitMHA(d_model, num_heads, device=device, dtype=dtype, operations=operations)
+        self.cross_attn_image = SplitMHA(d_model, num_heads, device=device, dtype=dtype, operations=operations)
+        self.linear1 = operations.Linear(d_model, dim_ff, device=device, dtype=dtype)
+        self.linear2 = operations.Linear(dim_ff, d_model, device=device, dtype=dtype)
+        self.norm1 = operations.LayerNorm(d_model, device=device, dtype=dtype)
+        self.norm2 = operations.LayerNorm(d_model, device=device, dtype=dtype)
+        self.norm3 = operations.LayerNorm(d_model, device=device, dtype=dtype)
+
+    def forward(self, x, pos, text_memory=None, text_mask=None):
+        normed = self.norm1(x)
+        q_k = normed + pos
+        x = x + self.self_attn(q_k, q_k, normed)
+        if text_memory is not None:
+            normed = self.norm2(x)
+            x = x + self.cross_attn_image(normed, text_memory, text_memory, mask=text_mask)
+        normed = self.norm3(x)
+        x = x + self.linear2(F.relu(self.linear1(normed)))
+        return x
+
+
+class TransformerEncoder(nn.Module):
+    """Checkpoint: transformer.encoder.layers.N.*"""
+    def __init__(self, d_model=256, num_heads=8, dim_ff=2048, num_layers=6, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.layers = nn.ModuleList([
+            EncoderLayer(d_model, num_heads, dim_ff, device=device, dtype=dtype, operations=operations)
+            for _ in range(num_layers)
+        ])
+
+    def forward(self, x, pos, text_memory=None, text_mask=None):
+        for layer in self.layers:
+            x = layer(x, pos, text_memory, text_mask)
+        return x
+
+
+class DecoderLayer(nn.Module):
+    def __init__(self, d_model=256, num_heads=8, dim_ff=2048, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.self_attn = SplitMHA(d_model, num_heads, device=device, dtype=dtype, operations=operations)
+        self.cross_attn = SplitMHA(d_model, num_heads, device=device, dtype=dtype, operations=operations)
+        self.ca_text = SplitMHA(d_model, num_heads, device=device, dtype=dtype, operations=operations)
+        self.norm1 = operations.LayerNorm(d_model, device=device, dtype=dtype)
+        self.norm2 = operations.LayerNorm(d_model, device=device, dtype=dtype)
+        self.norm3 = operations.LayerNorm(d_model, device=device, dtype=dtype)
+        self.catext_norm = operations.LayerNorm(d_model, device=device, dtype=dtype)
+        self.linear1 = operations.Linear(d_model, dim_ff, device=device, dtype=dtype)
+        self.linear2 = operations.Linear(dim_ff, d_model, device=device, dtype=dtype)
+
+    def forward(self, x, memory, x_pos, memory_pos, text_memory=None, text_mask=None, cross_attn_bias=None):
+        q_k = x + x_pos
+        x = self.norm2(x + self.self_attn(q_k, q_k, x))
+        if text_memory is not None:
+            x = self.catext_norm(x + self.ca_text(x + x_pos, text_memory, text_memory, mask=text_mask))
+        x = self.norm1(x + self.cross_attn(x + x_pos, memory + memory_pos, memory, mask=cross_attn_bias))
+        x = self.norm3(x + self.linear2(F.relu(self.linear1(x))))
+        return x
+
+
+class TransformerDecoder(nn.Module):
+    def __init__(self, d_model=256, num_heads=8, dim_ff=2048, num_layers=6,
+                 num_queries=200, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.d_model = d_model
+        self.num_queries = num_queries
+
+        self.layers = nn.ModuleList([
+            DecoderLayer(d_model, num_heads, dim_ff, device=device, dtype=dtype, operations=operations)
+            for _ in range(num_layers)
+        ])
+        self.norm = operations.LayerNorm(d_model, device=device, dtype=dtype)
+        self.query_embed = operations.Embedding(num_queries, d_model, device=device, dtype=dtype)
+        self.reference_points = operations.Embedding(num_queries, 4, device=device, dtype=dtype) # Reference points: Embedding(num_queries, 4) — learned anchor boxes
+        self.ref_point_head = MLP(d_model * 2, d_model, d_model, 2, device=device, dtype=dtype, operations=operations) # ref_point_head input: 512 (4 coords * 128 sine features each)
+        self.bbox_embed = MLP(d_model, d_model, 4, 3, device=device, dtype=dtype, operations=operations)
+
+        self.boxRPB_embed_x = MLP(2, d_model, num_heads, 2, device=device, dtype=dtype, operations=operations)
+        self.boxRPB_embed_y = MLP(2, d_model, num_heads, 2, device=device, dtype=dtype, operations=operations)
+
+        self.presence_token = operations.Embedding(1, d_model, device=device, dtype=dtype)
+        self.presence_token_head = MLP(d_model, d_model, 1, 3, device=device, dtype=dtype, operations=operations)
+        self.presence_token_out_norm = operations.LayerNorm(d_model, device=device, dtype=dtype)
+
+    @staticmethod
+    def _inverse_sigmoid(x):
+        return torch.log(x / (1 - x + 1e-6) + 1e-6)
+
+    def _compute_box_rpb(self, ref_points, H, W):
+        """Box rotary position bias: (B, Q, 4) cxcywh -> (B, n_heads, Q+1, H*W) bias."""
+        boxes_xyxy = box_cxcywh_to_xyxy(ref_points)
+        B, Q, _ = boxes_xyxy.shape
+        coords_h = torch.arange(H, device=ref_points.device, dtype=torch.float32) / H
+        coords_w = torch.arange(W, device=ref_points.device, dtype=torch.float32) / W
+        deltas_x = coords_w.view(1, 1, -1, 1) - boxes_xyxy[:, :, None, 0:3:2]
+        deltas_y = coords_h.view(1, 1, -1, 1) - boxes_xyxy[:, :, None, 1:4:2]
+
+        log2_8 = float(math.log2(8))
+        def log_scale(d):
+            return torch.sign(d * 8) * torch.log2(torch.abs(d * 8) + 1.0) / log2_8
+
+        rpb_x = self.boxRPB_embed_x(log_scale(deltas_x).to(ref_points.dtype))
+        rpb_y = self.boxRPB_embed_y(log_scale(deltas_y).to(ref_points.dtype))
+
+        bias = (rpb_y.unsqueeze(3) + rpb_x.unsqueeze(2)).flatten(2, 3).permute(0, 3, 1, 2)
+        pres_bias = torch.zeros(B, bias.shape[1], 1, bias.shape[3], device=bias.device, dtype=bias.dtype)
+        return torch.cat([pres_bias, bias], dim=2)
+
+    def forward(self, memory, memory_pos, text_memory=None, text_mask=None, H=72, W=72):
+        B = memory.shape[0]
+        tgt = cast_to_input(self.query_embed.weight, memory).unsqueeze(0).expand(B, -1, -1)
+        presence_out = cast_to_input(self.presence_token.weight, memory)[None].expand(B, -1, -1)
+        ref_points = cast_to_input(self.reference_points.weight, memory).unsqueeze(0).expand(B, -1, -1).sigmoid()
+
+        for layer_idx, layer in enumerate(self.layers):
+            query_pos = self.ref_point_head(gen_sineembed_for_position(ref_points, self.d_model))
+            tgt_with_pres = torch.cat([presence_out, tgt], dim=1)
+            pos_with_pres = torch.cat([torch.zeros_like(presence_out), query_pos], dim=1)
+            tgt_with_pres = layer(tgt_with_pres, memory, pos_with_pres, memory_pos,
+                                  text_memory, text_mask, self._compute_box_rpb(ref_points, H, W))
+            presence_out, tgt = tgt_with_pres[:, :1], tgt_with_pres[:, 1:]
+            if layer_idx < len(self.layers) - 1:
+                ref_inv = self._inverse_sigmoid(ref_points)
+                ref_points = (ref_inv + self.bbox_embed(self.norm(tgt))).sigmoid().detach()
+
+        query_out = self.norm(tgt)
+        ref_inv = self._inverse_sigmoid(ref_points)
+        boxes = (ref_inv + self.bbox_embed(query_out)).sigmoid()
+        presence = self.presence_token_head(self.presence_token_out_norm(presence_out)).squeeze(-1)
+        return {"decoder_output": query_out, "pred_boxes": boxes, "presence": presence}
+
+
+class Transformer(nn.Module):
+    def __init__(self, d_model=256, num_heads=8, dim_ff=2048, enc_layers=6, dec_layers=6,
+                 num_queries=200, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.encoder = TransformerEncoder(d_model, num_heads, dim_ff, enc_layers, device=device, dtype=dtype, operations=operations)
+        self.decoder = TransformerDecoder(d_model, num_heads, dim_ff, dec_layers, num_queries, device=device, dtype=dtype, operations=operations)
+
+
+class GeometryEncoder(nn.Module):
+    def __init__(self, d_model=256, num_heads=8, num_layers=3, roi_size=7, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.d_model = d_model
+        self.roi_size = roi_size
+        self.pos_enc = PositionEmbeddingSine(num_pos_feats=d_model, normalize=True)
+        self.points_direct_project = operations.Linear(2, d_model, device=device, dtype=dtype)
+        self.points_pool_project = operations.Linear(d_model, d_model, device=device, dtype=dtype)
+        self.points_pos_enc_project = operations.Linear(d_model, d_model, device=device, dtype=dtype)
+        self.boxes_direct_project = operations.Linear(4, d_model, device=device, dtype=dtype)
+        self.boxes_pool_project = operations.Conv2d(d_model, d_model, kernel_size=roi_size, device=device, dtype=dtype)
+        self.boxes_pos_enc_project = operations.Linear(d_model + 2, d_model, device=device, dtype=dtype)
+        self.label_embed = operations.Embedding(2, d_model, device=device, dtype=dtype)
+        self.cls_embed = operations.Embedding(1, d_model, device=device, dtype=dtype)
+        self.norm = operations.LayerNorm(d_model, device=device, dtype=dtype)
+        self.img_pre_norm = operations.LayerNorm(d_model, device=device, dtype=dtype)
+        self.encode = nn.ModuleList([
+            EncoderLayer(d_model, num_heads, 2048, device=device, dtype=dtype, operations=operations)
+            for _ in range(num_layers)
+        ])
+        self.encode_norm = operations.LayerNorm(d_model, device=device, dtype=dtype)
+        self.final_proj = operations.Linear(d_model, d_model, device=device, dtype=dtype)
+
+    def _encode_points(self, coords, labels, img_feat_2d):
+        """Encode point prompts: direct + pool + pos_enc + label. coords: [B, N, 2] normalized."""
+        B, N, _ = coords.shape
+        embed = self.points_direct_project(coords)
+        # Pool features from backbone at point locations via grid_sample
+        grid = (coords * 2 - 1).unsqueeze(2)  # [B, N, 1, 2] in [-1, 1]
+        sampled = F.grid_sample(img_feat_2d, grid, align_corners=False)  # [B, C, N, 1]
+        embed = embed + self.points_pool_project(sampled.squeeze(-1).permute(0, 2, 1))  # [B, N, C]
+        # Positional encoding of coordinates
+        x, y = coords[:, :, 0], coords[:, :, 1]  # [B, N]
+        pos_x, pos_y = self.pos_enc._encode_xy(x.flatten(), y.flatten())
+        enc = torch.cat([pos_x, pos_y], dim=-1).view(B, N, -1)
+        embed = embed + self.points_pos_enc_project(cast_to_input(enc, embed))
+        embed = embed + cast_to_input(self.label_embed(labels.long()), embed)
+        return embed
+
+    def _encode_boxes(self, boxes, labels, img_feat_2d):
+        """Encode box prompts: direct + pool + pos_enc + label. boxes: [B, N, 4] normalized cxcywh."""
+        B, N, _ = boxes.shape
+        embed = self.boxes_direct_project(boxes)
+        # ROI align from backbone at box regions
+        H, W = img_feat_2d.shape[-2:]
+        boxes_xyxy = box_cxcywh_to_xyxy(boxes)
+        scale = torch.tensor([W, H, W, H], dtype=boxes_xyxy.dtype, device=boxes_xyxy.device)
+        boxes_scaled = boxes_xyxy * scale
+        sampled = roi_align(img_feat_2d, boxes_scaled.view(-1, 4).split(N), self.roi_size)
+        proj = self.boxes_pool_project(sampled).view(B, N, -1)  # Conv2d(roi_size) -> [B*N, C, 1, 1] -> [B, N, C]
+        embed = embed + proj
+        # Positional encoding of box center + size
+        cx, cy, w, h = boxes[:, :, 0], boxes[:, :, 1], boxes[:, :, 2], boxes[:, :, 3]
+        enc = self.pos_enc.encode_boxes(cx.flatten(), cy.flatten(), w.flatten(), h.flatten())
+        enc = enc.view(B, N, -1)
+        embed = embed + self.boxes_pos_enc_project(cast_to_input(enc, embed))
+        embed = embed + cast_to_input(self.label_embed(labels.long()), embed)
+        return embed
+
+    def forward(self, points=None, boxes=None, image_features=None):
+        """Encode geometry prompts. image_features: [B, HW, C] flattened backbone features."""
+        # Prepare 2D image features for pooling
+        img_feat_2d = None
+        if image_features is not None:
+            B = image_features.shape[0]
+            HW, C = image_features.shape[1], image_features.shape[2]
+            hw = int(math.sqrt(HW))
+            img_normed = self.img_pre_norm(image_features)
+            img_feat_2d = img_normed.permute(0, 2, 1).view(B, C, hw, hw)
+
+        embeddings = []
+        if points is not None:
+            coords, labels = points
+            embeddings.append(self._encode_points(coords, labels, img_feat_2d))
+        if boxes is not None:
+            B = boxes.shape[0]
+            box_labels = torch.ones(B, boxes.shape[1], dtype=torch.long, device=boxes.device)
+            embeddings.append(self._encode_boxes(boxes, box_labels, img_feat_2d))
+        if not embeddings:
+            return None
+        geo = torch.cat(embeddings, dim=1)
+        geo = self.norm(geo)
+        if image_features is not None:
+            for layer in self.encode:
+                geo = layer(geo, torch.zeros_like(geo), image_features)
+        geo = self.encode_norm(geo)
+        return self.final_proj(geo)
+
+
+class PixelDecoder(nn.Module):
+    """Top-down FPN pixel decoder with GroupNorm + ReLU + nearest interpolation."""
+    def __init__(self, d_model=256, num_stages=3, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.conv_layers = nn.ModuleList([operations.Conv2d(d_model, d_model, kernel_size=3, padding=1, device=device, dtype=dtype) for _ in range(num_stages)])
+        self.norms = nn.ModuleList([operations.GroupNorm(8, d_model, device=device, dtype=dtype) for _ in range(num_stages)])
+
+    def forward(self, backbone_features):
+        prev = backbone_features[-1]
+        for i, feat in enumerate(backbone_features[:-1][::-1]):
+            prev = F.relu(self.norms[i](self.conv_layers[i](feat + F.interpolate(prev, size=feat.shape[-2:], mode="nearest"))))
+        return prev
+
+
+class MaskPredictor(nn.Module):
+    def __init__(self, d_model=256, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.mask_embed = MLP(d_model, d_model, d_model, 3, device=device, dtype=dtype, operations=operations)
+
+    def forward(self, query_embeddings, pixel_features):
+        mask_embed = self.mask_embed(query_embeddings)
+        return torch.einsum("bqc,bchw->bqhw", mask_embed, pixel_features)
+
+
+class SegmentationHead(nn.Module):
+    def __init__(self, d_model=256, num_heads=8, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.d_model = d_model
+        self.pixel_decoder = PixelDecoder(d_model, 3, device=device, dtype=dtype, operations=operations)
+        self.mask_predictor = MaskPredictor(d_model, device=device, dtype=dtype, operations=operations)
+        self.cross_attend_prompt = SplitMHA(d_model, num_heads, device=device, dtype=dtype, operations=operations)
+        self.cross_attn_norm = operations.LayerNorm(d_model, device=device, dtype=dtype)
+        self.instance_seg_head = operations.Conv2d(d_model, d_model, kernel_size=1, device=device, dtype=dtype)
+        self.semantic_seg_head = operations.Conv2d(d_model, 1, kernel_size=1, device=device, dtype=dtype)
+
+    def forward(self, query_embeddings, backbone_features, encoder_hidden_states=None, prompt=None, prompt_mask=None):
+        if encoder_hidden_states is not None and prompt is not None:
+            enc_normed = self.cross_attn_norm(encoder_hidden_states)
+            enc_cross = self.cross_attend_prompt(enc_normed, prompt, prompt, mask=prompt_mask)
+            encoder_hidden_states = enc_cross + encoder_hidden_states
+
+        if encoder_hidden_states is not None:
+            B, H, W = encoder_hidden_states.shape[0], backbone_features[-1].shape[-2], backbone_features[-1].shape[-1]
+            encoder_visual = encoder_hidden_states[:, :H * W].permute(0, 2, 1).view(B, self.d_model, H, W)
+            backbone_features = list(backbone_features)
+            backbone_features[-1] = encoder_visual
+
+        pixel_features = self.pixel_decoder(backbone_features)
+        instance_features = self.instance_seg_head(pixel_features)
+        masks = self.mask_predictor(query_embeddings, instance_features)
+        return masks
+
+
+class DotProductScoring(nn.Module):
+    def __init__(self, d_model=256, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.hs_proj = operations.Linear(d_model, d_model, device=device, dtype=dtype)
+        self.prompt_proj = operations.Linear(d_model, d_model, device=device, dtype=dtype)
+        self.prompt_mlp = MLPWithNorm(d_model, 2048, d_model, 2, device=device, dtype=dtype, operations=operations)
+        self.scale = 1.0 / (d_model ** 0.5)
+
+    def forward(self, query_embeddings, prompt_embeddings, prompt_mask=None):
+        prompt = self.prompt_mlp(prompt_embeddings)
+        if prompt_mask is not None:
+            weight = prompt_mask.unsqueeze(-1).to(dtype=prompt.dtype)
+            pooled = (prompt * weight).sum(dim=1) / weight.sum(dim=1).clamp(min=1)
+        else:
+            pooled = prompt.mean(dim=1)
+        hs = self.hs_proj(query_embeddings)
+        pp = self.prompt_proj(pooled).unsqueeze(-1).to(hs.dtype)
+        scores = torch.matmul(hs, pp)
+        return (scores * self.scale).clamp(-12.0, 12.0).squeeze(-1)
+
+
+class SAM3Detector(nn.Module):
+    def __init__(self, d_model=256, embed_dim=1024, num_queries=200, device=None, dtype=None, operations=None, **kwargs):
+        super().__init__()
+        image_model = kwargs.pop("image_model", "SAM3")
+        for k in ("num_heads", "num_head_channels"):
+            kwargs.pop(k, None)
+        multiplex = image_model == "SAM31"
+        # SAM3: 4 FPN levels, drop last (scalp=1); SAM3.1: 3 levels, use all (scalp=0)
+        self.scalp = 0 if multiplex else 1
+        self.backbone = nn.ModuleDict({
+            "vision_backbone": SAM3VisionBackbone(embed_dim=embed_dim, d_model=d_model, multiplex=multiplex, device=device, dtype=dtype, operations=operations, **kwargs),
+            "language_backbone": nn.ModuleDict({"resizer": operations.Linear(embed_dim, d_model, device=device, dtype=dtype)}),
+        })
+        self.transformer = Transformer(d_model=d_model, num_queries=num_queries, device=device, dtype=dtype, operations=operations)
+        self.segmentation_head = SegmentationHead(d_model=d_model, device=device, dtype=dtype, operations=operations)
+        self.geometry_encoder = GeometryEncoder(d_model=d_model, device=device, dtype=dtype, operations=operations)
+        self.dot_prod_scoring = DotProductScoring(d_model=d_model, device=device, dtype=dtype, operations=operations)
+
+    def _get_backbone_features(self, images):
+        """Run backbone and return (detector_features, detector_positions, tracker_features, tracker_positions)."""
+        bb = self.backbone["vision_backbone"]
+        if bb.multiplex:
+            all_f, all_p, tf, tp = bb(images, tracker_mode="propagation")
+        else:
+            all_f, all_p, tf, tp = bb(images, need_tracker=True)
+        return all_f, all_p, tf, tp
+
+    @staticmethod
+    def _run_geo_layer(layer, x, memory, memory_pos):
+        x = x + layer.self_attn(layer.norm1(x))
+        x = x + layer.cross_attn_image(layer.norm2(x), memory + memory_pos, memory)
+        x = x + layer.linear2(F.relu(layer.linear1(layer.norm3(x))))
+        return x
+
+    def _detect(self, features, positions, text_embeddings=None, text_mask=None,
+                points=None, boxes=None):
+        """Shared detection: geometry encoding, transformer, scoring, segmentation."""
+        B = features[0].shape[0]
+        # Scalp for encoder (use top-level feature), but keep all levels for segmentation head
+        seg_features = features
+        if self.scalp > 0:
+            features = features[:-self.scalp]
+            positions = positions[:-self.scalp]
+        enc_feat, enc_pos = features[-1], positions[-1]
+        _, _, H, W = enc_feat.shape
+        img_flat = enc_feat.flatten(2).permute(0, 2, 1)
+        pos_flat = enc_pos.flatten(2).permute(0, 2, 1)
+
+        has_prompts = text_embeddings is not None or points is not None or boxes is not None
+        if has_prompts:
+            geo_enc = self.geometry_encoder
+            geo_prompts = geo_enc(points=points, boxes=boxes, image_features=img_flat)
+            geo_cls = geo_enc.norm(geo_enc.final_proj(cast_to_input(geo_enc.cls_embed.weight, img_flat).view(1, 1, -1).expand(B, -1, -1)))
+            for layer in geo_enc.encode:
+                geo_cls = self._run_geo_layer(layer, geo_cls, img_flat, pos_flat)
+            geo_cls = geo_enc.encode_norm(geo_cls)
+            if text_embeddings is not None and text_embeddings.shape[0] != B:
+                text_embeddings = text_embeddings.expand(B, -1, -1)
+            if text_mask is not None and text_mask.shape[0] != B:
+                text_mask = text_mask.expand(B, -1)
+            parts = [t for t in [text_embeddings, geo_prompts, geo_cls] if t is not None]
+            text_embeddings = torch.cat(parts, dim=1)
+            n_new = text_embeddings.shape[1] - (text_mask.shape[1] if text_mask is not None else 0)
+            if text_mask is not None:
+                text_mask = torch.cat([text_mask, torch.ones(B, n_new, dtype=torch.bool, device=text_mask.device)], dim=1)
+            else:
+                text_mask = torch.ones(B, text_embeddings.shape[1], dtype=torch.bool, device=text_embeddings.device)
+
+        memory = self.transformer.encoder(img_flat, pos_flat, text_embeddings, text_mask)
+        dec_out = self.transformer.decoder(memory, pos_flat, text_embeddings, text_mask, H, W)
+        query_out, pred_boxes = dec_out["decoder_output"], dec_out["pred_boxes"]
+
+        if text_embeddings is not None:
+            scores = self.dot_prod_scoring(query_out, text_embeddings, text_mask)
+        else:
+            scores = torch.zeros(B, query_out.shape[1], device=query_out.device)
+
+        masks = self.segmentation_head(query_out, seg_features, encoder_hidden_states=memory, prompt=text_embeddings, prompt_mask=text_mask)
+        return box_cxcywh_to_xyxy(pred_boxes), scores, masks, dec_out
+
+    def forward(self, images, text_embeddings=None, text_mask=None, points=None, boxes=None, threshold=0.3, orig_size=None):
+        features, positions, _, _ = self._get_backbone_features(images)
+
+        if text_embeddings is not None:
+            text_embeddings = self.backbone["language_backbone"]["resizer"](text_embeddings)
+            if text_mask is not None:
+                text_mask = text_mask.bool()
+
+        boxes_xyxy, scores, masks, dec_out = self._detect(
+            features, positions, text_embeddings, text_mask, points, boxes)
+
+        if orig_size is not None:
+            oh, ow = orig_size
+            boxes_xyxy = boxes_xyxy * torch.tensor([ow, oh, ow, oh], device=boxes_xyxy.device, dtype=boxes_xyxy.dtype)
+            masks = F.interpolate(masks, size=orig_size, mode="bilinear", align_corners=False)
+
+        return {
+            "boxes": boxes_xyxy,
+            "scores": scores,
+            "masks": masks,
+            "presence": dec_out.get("presence"),
+        }
+
+    def forward_from_trunk(self, trunk_out, text_embeddings, text_mask):
+        """Run detection using a pre-computed ViTDet trunk output.
+
+        text_embeddings must already be resized through language_backbone.resizer.
+        Returns dict with boxes (normalized xyxy), scores, masks at detector resolution.
+        """
+        bb = self.backbone["vision_backbone"]
+        features = [conv(trunk_out) for conv in bb.convs]
+        positions = [cast_to_input(bb.position_encoding(f), f) for f in features]
+
+        if text_mask is not None:
+            text_mask = text_mask.bool()
+
+        boxes_xyxy, scores, masks, _ = self._detect(features, positions, text_embeddings, text_mask)
+        return {"boxes": boxes_xyxy, "scores": scores, "masks": masks}
+
+
+class SAM3Model(nn.Module):
+    def __init__(self, device=None, dtype=None, operations=None, **kwargs):
+        super().__init__()
+        self.dtype = dtype
+        image_model = kwargs.get("image_model", "SAM3")
+        tracker_cls = TRACKER_CLASSES[image_model]
+        self.detector = SAM3Detector(device=device, dtype=dtype, operations=operations, **kwargs)
+        self.tracker = tracker_cls(device=device, dtype=dtype, operations=operations, **kwargs)
+
+    def forward(self, images, **kwargs):
+        return self.detector(images, **kwargs)
+
+    def forward_segment(self, images, point_inputs=None, box_inputs=None, mask_inputs=None):
+        """Interactive segmentation using SAM decoder with point/box/mask prompts.
+
+        Args:
+            images: [B, 3, 1008, 1008] preprocessed images
+            point_inputs: {"point_coords": [B, N, 2], "point_labels": [B, N]} in 1008x1008 pixel space
+            box_inputs: [B, 2, 2] box corners (top-left, bottom-right) in 1008x1008 pixel space
+            mask_inputs: [B, 1, H, W] coarse mask logits to refine
+        Returns:
+            [B, 1, image_size, image_size] high-res mask logits
+        """
+        bb = self.detector.backbone["vision_backbone"]
+        if bb.multiplex:
+            _, _, tracker_features, tracker_positions = bb(images, tracker_mode="interactive")
+        else:
+            _, _, tracker_features, tracker_positions = bb(images, need_tracker=True)
+            if self.detector.scalp > 0:
+                tracker_features = tracker_features[:-self.detector.scalp]
+                tracker_positions = tracker_positions[:-self.detector.scalp]
+
+        high_res = list(tracker_features[:-1])
+        backbone_feat = tracker_features[-1]
+        B, C, H, W = backbone_feat.shape
+        # Add no-memory embedding (init frame path)
+        no_mem = getattr(self.tracker, 'interactivity_no_mem_embed', None)
+        if no_mem is None:
+            no_mem = getattr(self.tracker, 'no_mem_embed', None)
+        if no_mem is not None:
+            feat_flat = backbone_feat.flatten(2).permute(0, 2, 1)
+            feat_flat = feat_flat + cast_to_input(no_mem, feat_flat)
+            backbone_feat = feat_flat.view(B, H, W, C).permute(0, 3, 1, 2)
+
+        num_pts = 0 if point_inputs is None else point_inputs["point_labels"].size(1)
+        _, high_res_masks, _, _ = self.tracker._forward_sam_heads(
+            backbone_features=backbone_feat,
+            point_inputs=point_inputs,
+            mask_inputs=mask_inputs,
+            box_inputs=box_inputs,
+            high_res_features=high_res,
+            multimask_output=(0 < num_pts <= 1),
+        )
+        return high_res_masks
+
+    def forward_video(self, images, initial_masks, pbar=None, text_prompts=None,
+                       new_det_thresh=0.5, max_objects=0, detect_interval=1):
+        """Track video with optional per-frame text-prompted detection."""
+        bb = self.detector.backbone["vision_backbone"]
+
+        def backbone_fn(frame, frame_idx=None):
+            trunk_out = bb.trunk(frame)
+            if bb.multiplex:
+                _, _, tf, tp = bb(frame, tracker_mode="propagation", cached_trunk=trunk_out, tracker_only=True)
+            else:
+                _, _, tf, tp = bb(frame, need_tracker=True, cached_trunk=trunk_out, tracker_only=True)
+            return tf, tp, trunk_out
+
+        detect_fn = None
+        if text_prompts:
+            resizer = self.detector.backbone["language_backbone"]["resizer"]
+            resized = [(resizer(emb), m.bool() if m is not None else None) for emb, m in text_prompts]
+            def detect_fn(trunk_out):
+                all_scores, all_masks = [], []
+                for emb, mask in resized:
+                    det = self.detector.forward_from_trunk(trunk_out, emb, mask)
+                    all_scores.append(det["scores"])
+                    all_masks.append(det["masks"])
+                return {"scores": torch.cat(all_scores, dim=1), "masks": torch.cat(all_masks, dim=1)}
+
+        if hasattr(self.tracker, 'track_video_with_detection'):
+            return self.tracker.track_video_with_detection(
+                backbone_fn, images, initial_masks, detect_fn,
+                new_det_thresh=new_det_thresh, max_objects=max_objects,
+                detect_interval=detect_interval, backbone_obj=bb, pbar=pbar)
+        # SAM3 (non-multiplex) — no detection support, requires initial masks
+        if initial_masks is None:
+            raise ValueError("SAM3 (non-multiplex) requires initial_mask for video tracking")
+        return self.tracker.track_video(backbone_fn, images, initial_masks, pbar=pbar, backbone_obj=bb)

comfy/ldm/sam3/sam.py

@@ -0,0 +1,425 @@
+# SAM3 shared components: primitives, ViTDet backbone, FPN neck, position encodings.
+
+import math
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from comfy.ldm.modules.attention import optimized_attention
+from comfy.ldm.flux.math import apply_rope
+from comfy.ldm.flux.layers import EmbedND
+from comfy.ops import cast_to_input
+
+
+class MLP(nn.Module):
+    def __init__(self, input_dim, hidden_dim, output_dim, num_layers, sigmoid_output=False, device=None, dtype=None, operations=None):
+        super().__init__()
+        dims = [input_dim] + [hidden_dim] * (num_layers - 1) + [output_dim]
+        self.layers = nn.ModuleList([operations.Linear(dims[i], dims[i + 1], device=device, dtype=dtype) for i in range(num_layers)])
+        self.sigmoid_output = sigmoid_output
+
+    def forward(self, x):
+        for i, layer in enumerate(self.layers):
+            x = F.relu(layer(x)) if i < len(self.layers) - 1 else layer(x)
+        return torch.sigmoid(x) if self.sigmoid_output else x
+
+
+class SAMAttention(nn.Module):
+    def __init__(self, embedding_dim, num_heads, downsample_rate=1, kv_in_dim=None, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.num_heads = num_heads
+        internal_dim = embedding_dim // downsample_rate
+        kv_dim = kv_in_dim if kv_in_dim is not None else embedding_dim
+        self.q_proj = operations.Linear(embedding_dim, internal_dim, device=device, dtype=dtype)
+        self.k_proj = operations.Linear(kv_dim, internal_dim, device=device, dtype=dtype)
+        self.v_proj = operations.Linear(kv_dim, internal_dim, device=device, dtype=dtype)
+        self.out_proj = operations.Linear(internal_dim, embedding_dim, device=device, dtype=dtype)
+
+    def forward(self, q, k, v):
+        q = self.q_proj(q)
+        k = self.k_proj(k)
+        v = self.v_proj(v)
+        return self.out_proj(optimized_attention(q, k, v, self.num_heads, low_precision_attention=False))
+
+
+class TwoWayAttentionBlock(nn.Module):
+    def __init__(self, embedding_dim, num_heads, mlp_dim=2048, attention_downsample_rate=2, skip_first_layer_pe=False, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.skip_first_layer_pe = skip_first_layer_pe
+        self.self_attn = SAMAttention(embedding_dim, num_heads, device=device, dtype=dtype, operations=operations)
+        self.cross_attn_token_to_image = SAMAttention(embedding_dim, num_heads, downsample_rate=attention_downsample_rate, device=device, dtype=dtype, operations=operations)
+        self.cross_attn_image_to_token = SAMAttention(embedding_dim, num_heads, downsample_rate=attention_downsample_rate, device=device, dtype=dtype, operations=operations)
+        self.mlp = nn.Sequential(operations.Linear(embedding_dim, mlp_dim, device=device, dtype=dtype), nn.ReLU(), operations.Linear(mlp_dim, embedding_dim, device=device, dtype=dtype))
+        self.norm1 = operations.LayerNorm(embedding_dim, device=device, dtype=dtype)
+        self.norm2 = operations.LayerNorm(embedding_dim, device=device, dtype=dtype)
+        self.norm3 = operations.LayerNorm(embedding_dim, device=device, dtype=dtype)
+        self.norm4 = operations.LayerNorm(embedding_dim, device=device, dtype=dtype)
+
+    def forward(self, queries, keys, query_pe, key_pe):
+        if self.skip_first_layer_pe:
+            queries = self.norm1(self.self_attn(queries, queries, queries))
+        else:
+            q = queries + query_pe
+            queries = self.norm1(queries + self.self_attn(q, q, queries))
+        q, k = queries + query_pe, keys + key_pe
+        queries = self.norm2(queries + self.cross_attn_token_to_image(q, k, keys))
+        queries = self.norm3(queries + self.mlp(queries))
+        q, k = queries + query_pe, keys + key_pe
+        keys = self.norm4(keys + self.cross_attn_image_to_token(k, q, queries))
+        return queries, keys
+
+
+class TwoWayTransformer(nn.Module):
+    def __init__(self, depth=2, embedding_dim=256, num_heads=8, mlp_dim=2048, attention_downsample_rate=2, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.layers = nn.ModuleList([
+            TwoWayAttentionBlock(embedding_dim, num_heads, mlp_dim, attention_downsample_rate,
+                                 skip_first_layer_pe=(i == 0), device=device, dtype=dtype, operations=operations)
+            for i in range(depth)
+        ])
+        self.final_attn_token_to_image = SAMAttention(embedding_dim, num_heads, downsample_rate=attention_downsample_rate, device=device, dtype=dtype, operations=operations)
+        self.norm_final = operations.LayerNorm(embedding_dim, device=device, dtype=dtype)
+
+    def forward(self, image_embedding, image_pe, point_embedding):
+        queries, keys = point_embedding, image_embedding
+        for layer in self.layers:
+            queries, keys = layer(queries, keys, point_embedding, image_pe)
+        q, k = queries + point_embedding, keys + image_pe
+        queries = self.norm_final(queries + self.final_attn_token_to_image(q, k, keys))
+        return queries, keys
+
+
+class PositionEmbeddingRandom(nn.Module):
+    """Fourier feature positional encoding with random gaussian projection."""
+    def __init__(self, num_pos_feats=64, scale=None):
+        super().__init__()
+        self.register_buffer("positional_encoding_gaussian_matrix", (scale or 1.0) * torch.randn(2, num_pos_feats))
+
+    def _encode(self, normalized_coords):
+        """Map normalized [0,1] coordinates to fourier features via random projection. Computes in fp32."""
+        orig_dtype = normalized_coords.dtype
+        proj_matrix = self.positional_encoding_gaussian_matrix.to(device=normalized_coords.device, dtype=torch.float32)
+        projected = 2 * math.pi * (2 * normalized_coords.float() - 1) @ proj_matrix
+        return torch.cat([projected.sin(), projected.cos()], dim=-1).to(orig_dtype)
+
+    def forward(self, size, device=None):
+        h, w = size
+        dev = device if device is not None else self.positional_encoding_gaussian_matrix.device
+        ones = torch.ones((h, w), device=dev, dtype=torch.float32)
+        norm_xy = torch.stack([(ones.cumsum(1) - 0.5) / w, (ones.cumsum(0) - 0.5) / h], dim=-1)
+        return self._encode(norm_xy).permute(2, 0, 1).unsqueeze(0)
+
+    def forward_with_coords(self, pixel_coords, image_size):
+        norm = pixel_coords.clone()
+        norm[:, :, 0] /= image_size[1]
+        norm[:, :, 1] /= image_size[0]
+        return self._encode(norm)
+
+
+# ViTDet backbone + FPN neck
+
+def window_partition(x: torch.Tensor, window_size: int):
+    B, H, W, C = x.shape
+    pad_h = (window_size - H % window_size) % window_size
+    pad_w = (window_size - W % window_size) % window_size
+    if pad_h > 0 or pad_w > 0:
+        x = F.pad(x, (0, 0, 0, pad_w, 0, pad_h))
+    Hp, Wp = H + pad_h, W + pad_w
+    x = x.view(B, Hp // window_size, window_size, Wp // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows, (Hp, Wp)
+
+
+def window_unpartition(windows: torch.Tensor, window_size: int, pad_hw, hw):
+    Hp, Wp = pad_hw
+    H, W = hw
+    B = windows.shape[0] // (Hp * Wp // window_size // window_size)
+    x = windows.view(B, Hp // window_size, Wp // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, Hp, Wp, -1)
+    if Hp > H or Wp > W:
+        x = x[:, :H, :W, :].contiguous()
+    return x
+
+
+def rope_2d(end_x: int, end_y: int, dim: int, theta: float = 10000.0, scale_pos: float = 1.0):
+    """Generate 2D axial RoPE using flux EmbedND. Returns [1, 1, HW, dim//2, 2, 2]."""
+    t = torch.arange(end_x * end_y, dtype=torch.float32)
+    ids = torch.stack([(t % end_x) * scale_pos,
+                       torch.div(t, end_x, rounding_mode="floor") * scale_pos], dim=-1)
+    return EmbedND(dim=dim, theta=theta, axes_dim=[dim // 2, dim // 2])(ids.unsqueeze(0))
+
+
+class _ViTMLP(nn.Module):
+    def __init__(self, dim, mlp_ratio=4.0, device=None, dtype=None, operations=None):
+        super().__init__()
+        hidden = int(dim * mlp_ratio)
+        self.fc1 = operations.Linear(dim, hidden, device=device, dtype=dtype)
+        self.act = nn.GELU()
+        self.fc2 = operations.Linear(hidden, dim, device=device, dtype=dtype)
+
+    def forward(self, x):
+        return self.fc2(self.act(self.fc1(x)))
+
+
+class Attention(nn.Module):
+    """ViTDet multi-head attention with fused QKV projection."""
+
+    def __init__(self, dim, num_heads=8, qkv_bias=True, use_rope=False, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.use_rope = use_rope
+        self.qkv = operations.Linear(dim, dim * 3, bias=qkv_bias, device=device, dtype=dtype)
+        self.proj = operations.Linear(dim, dim, device=device, dtype=dtype)
+
+    def forward(self, x, freqs_cis=None):
+        B, N, C = x.shape
+        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, self.head_dim)
+        q, k, v = qkv.permute(2, 0, 3, 1, 4).unbind(dim=0)
+        if self.use_rope and freqs_cis is not None:
+            q, k = apply_rope(q, k, freqs_cis)
+        return self.proj(optimized_attention(q, k, v, self.num_heads, skip_reshape=True, low_precision_attention=False))
+
+
+class Block(nn.Module):
+    def __init__(self, dim, num_heads, mlp_ratio=4.0, qkv_bias=True, window_size=0, use_rope=False, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.window_size = window_size
+        self.norm1 = operations.LayerNorm(dim, device=device, dtype=dtype)
+        self.attn = Attention(dim, num_heads, qkv_bias, use_rope, device=device, dtype=dtype, operations=operations)
+        self.norm2 = operations.LayerNorm(dim, device=device, dtype=dtype)
+        self.mlp = _ViTMLP(dim, mlp_ratio, device=device, dtype=dtype, operations=operations)
+
+    def forward(self, x, freqs_cis=None):
+        shortcut = x
+        x = self.norm1(x)
+        if self.window_size > 0:
+            H, W = x.shape[1], x.shape[2]
+            x, pad_hw = window_partition(x, self.window_size)
+            x = x.view(x.shape[0], self.window_size * self.window_size, -1)
+            x = self.attn(x, freqs_cis=freqs_cis)
+            x = x.view(-1, self.window_size, self.window_size, x.shape[-1])
+            x = window_unpartition(x, self.window_size, pad_hw, (H, W))
+        else:
+            B, H, W, C = x.shape
+            x = x.view(B, H * W, C)
+            x = self.attn(x, freqs_cis=freqs_cis)
+            x = x.view(B, H, W, C)
+        x = shortcut + x
+        x = x + self.mlp(self.norm2(x))
+        return x
+
+
+class PatchEmbed(nn.Module):
+    def __init__(self, patch_size=14, in_chans=3, embed_dim=1024, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.proj = operations.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size, bias=False, device=device, dtype=dtype)
+
+    def forward(self, x):
+        return self.proj(x)
+
+
+class ViTDet(nn.Module):
+    def __init__(self, img_size=1008, patch_size=14, embed_dim=1024, depth=32, num_heads=16, mlp_ratio=4.625, qkv_bias=True, window_size=24,
+                 global_att_blocks=(7, 15, 23, 31), use_rope=True, pretrain_img_size=336, device=None, dtype=None, operations=None, **kwargs):
+        super().__init__()
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.global_att_blocks = set(global_att_blocks)
+
+        self.patch_embed = PatchEmbed(patch_size, 3, embed_dim, device=device, dtype=dtype, operations=operations)
+
+        num_patches = (pretrain_img_size // patch_size) ** 2 + 1  # +1 for cls token
+        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim, device=device, dtype=dtype))
+
+        self.ln_pre = operations.LayerNorm(embed_dim, device=device, dtype=dtype)
+
+        grid_size = img_size // patch_size
+        pretrain_grid = pretrain_img_size // patch_size
+
+        self.blocks = nn.ModuleList()
+        for i in range(depth):
+            is_global = i in self.global_att_blocks
+            self.blocks.append(Block(
+                embed_dim, num_heads, mlp_ratio, qkv_bias,
+                window_size=0 if is_global else window_size,
+                use_rope=use_rope,
+                device=device, dtype=dtype, operations=operations,
+            ))
+
+        if use_rope:
+            rope_scale = pretrain_grid / grid_size
+            self.register_buffer("freqs_cis", rope_2d(grid_size, grid_size, embed_dim // num_heads, scale_pos=rope_scale), persistent=False)
+            self.register_buffer("freqs_cis_window", rope_2d(window_size, window_size, embed_dim // num_heads), persistent=False)
+        else:
+            self.freqs_cis = None
+            self.freqs_cis_window = None
+
+    def _get_pos_embed(self, num_tokens):
+        pos = self.pos_embed
+        if pos.shape[1] == num_tokens:
+            return pos
+        cls_pos = pos[:, :1]
+        spatial_pos = pos[:, 1:]
+        old_size = int(math.sqrt(spatial_pos.shape[1]))
+        new_size = int(math.sqrt(num_tokens - 1)) if num_tokens > 1 else old_size
+        spatial_2d = spatial_pos.reshape(1, old_size, old_size, -1).permute(0, 3, 1, 2)
+        tiles_h = new_size // old_size + 1
+        tiles_w = new_size // old_size + 1
+        tiled = spatial_2d.tile([1, 1, tiles_h, tiles_w])[:, :, :new_size, :new_size]
+        tiled = tiled.permute(0, 2, 3, 1).reshape(1, new_size * new_size, -1)
+        return torch.cat([cls_pos, tiled], dim=1)
+
+    def forward(self, x):
+        x = self.patch_embed(x)
+        B, C, Hp, Wp = x.shape
+        x = x.permute(0, 2, 3, 1).reshape(B, Hp * Wp, C)
+
+        pos = cast_to_input(self._get_pos_embed(Hp * Wp + 1), x)
+        x = x + pos[:, 1:Hp * Wp + 1]
+
+        x = x.view(B, Hp, Wp, C)
+        x = self.ln_pre(x)
+
+        freqs_cis_global = self.freqs_cis
+        freqs_cis_win = self.freqs_cis_window
+        if freqs_cis_global is not None:
+            freqs_cis_global = cast_to_input(freqs_cis_global, x)
+        if freqs_cis_win is not None:
+            freqs_cis_win = cast_to_input(freqs_cis_win, x)
+
+        for block in self.blocks:
+            fc = freqs_cis_win if block.window_size > 0 else freqs_cis_global
+            x = block(x, freqs_cis=fc)
+
+        return x.permute(0, 3, 1, 2)
+
+
+class FPNScaleConv(nn.Module):
+    def __init__(self, in_dim, out_dim, scale, device=None, dtype=None, operations=None):
+        super().__init__()
+        if scale == 4.0:
+            self.dconv_2x2_0 = operations.ConvTranspose2d(in_dim, in_dim // 2, kernel_size=2, stride=2, device=device, dtype=dtype)
+            self.dconv_2x2_1 = operations.ConvTranspose2d(in_dim // 2, in_dim // 4, kernel_size=2, stride=2, device=device, dtype=dtype)
+            proj_in = in_dim // 4
+        elif scale == 2.0:
+            self.dconv_2x2 = operations.ConvTranspose2d(in_dim, in_dim // 2, kernel_size=2, stride=2, device=device, dtype=dtype)
+            proj_in = in_dim // 2
+        elif scale == 1.0:
+            proj_in = in_dim
+        elif scale == 0.5:
+            self.pool = nn.MaxPool2d(kernel_size=2, stride=2)
+            proj_in = in_dim
+        self.scale = scale
+        self.conv_1x1 = operations.Conv2d(proj_in, out_dim, kernel_size=1, device=device, dtype=dtype)
+        self.conv_3x3 = operations.Conv2d(out_dim, out_dim, kernel_size=3, padding=1, device=device, dtype=dtype)
+
+    def forward(self, x):
+        if self.scale == 4.0:
+            x = F.gelu(self.dconv_2x2_0(x))
+            x = self.dconv_2x2_1(x)
+        elif self.scale == 2.0:
+            x = self.dconv_2x2(x)
+        elif self.scale == 0.5:
+            x = self.pool(x)
+        x = self.conv_1x1(x)
+        x = self.conv_3x3(x)
+        return x
+
+
+class PositionEmbeddingSine(nn.Module):
+    """2D sinusoidal position encoding (DETR-style) with result caching."""
+    def __init__(self, num_pos_feats=256, temperature=10000.0, normalize=True, scale=None):
+        super().__init__()
+        assert num_pos_feats % 2 == 0
+        self.half_dim = num_pos_feats // 2
+        self.temperature = temperature
+        self.normalize = normalize
+        self.scale = scale if scale is not None else 2 * math.pi
+        self._cache = {}
+
+    def _sincos(self, vals):
+        """Encode 1D values to interleaved sin/cos features."""
+        freqs = self.temperature ** (2 * (torch.arange(self.half_dim, dtype=torch.float32, device=vals.device) // 2) / self.half_dim)
+        raw = vals[..., None] * self.scale / freqs
+        return torch.stack((raw[..., 0::2].sin(), raw[..., 1::2].cos()), dim=-1).flatten(-2)
+
+    def _encode_xy(self, x, y):
+        """Encode normalized x, y coordinates to sinusoidal features. Returns (pos_x, pos_y) each [N, half_dim]."""
+        dim_t = self.temperature ** (2 * (torch.arange(self.half_dim, dtype=torch.float32, device=x.device) // 2) / self.half_dim)
+        pos_x = x[:, None] * self.scale / dim_t
+        pos_y = y[:, None] * self.scale / dim_t
+        pos_x = torch.stack((pos_x[:, 0::2].sin(), pos_x[:, 1::2].cos()), dim=2).flatten(1)
+        pos_y = torch.stack((pos_y[:, 0::2].sin(), pos_y[:, 1::2].cos()), dim=2).flatten(1)
+        return pos_x, pos_y
+
+    def encode_boxes(self, cx, cy, w, h):
+        """Encode box center + size to [N, d_model+2] features."""
+        pos_x, pos_y = self._encode_xy(cx, cy)
+        return torch.cat((pos_y, pos_x, h[:, None], w[:, None]), dim=1)
+
+    def forward(self, x):
+        B, C, H, W = x.shape
+        key = (H, W, x.device)
+        if key not in self._cache:
+            gy = torch.arange(H, dtype=torch.float32, device=x.device)
+            gx = torch.arange(W, dtype=torch.float32, device=x.device)
+            if self.normalize:
+                gy, gx = gy / (H - 1 + 1e-6), gx / (W - 1 + 1e-6)
+            yy, xx = torch.meshgrid(gy, gx, indexing="ij")
+            self._cache[key] = torch.cat((self._sincos(yy), self._sincos(xx)), dim=-1).permute(2, 0, 1).unsqueeze(0)
+        return self._cache[key].expand(B, -1, -1, -1)
+
+
+class SAM3VisionBackbone(nn.Module):
+    def __init__(self, embed_dim=1024, d_model=256, multiplex=False, device=None, dtype=None, operations=None, **kwargs):
+        super().__init__()
+        self.trunk = ViTDet(embed_dim=embed_dim, device=device, dtype=dtype, operations=operations, **kwargs)
+        self.position_encoding = PositionEmbeddingSine(num_pos_feats=d_model, normalize=True)
+        self.multiplex = multiplex
+
+        fpn_args = dict(device=device, dtype=dtype, operations=operations)
+        if multiplex:
+            scales = [4.0, 2.0, 1.0]
+            self.convs = nn.ModuleList([FPNScaleConv(embed_dim, d_model, s, **fpn_args) for s in scales])
+            self.propagation_convs = nn.ModuleList([FPNScaleConv(embed_dim, d_model, s, **fpn_args) for s in scales])
+            self.interactive_convs = nn.ModuleList([FPNScaleConv(embed_dim, d_model, s, **fpn_args) for s in scales])
+        else:
+            scales = [4.0, 2.0, 1.0, 0.5]
+            self.convs = nn.ModuleList([FPNScaleConv(embed_dim, d_model, s, **fpn_args) for s in scales])
+            self.sam2_convs = nn.ModuleList([FPNScaleConv(embed_dim, d_model, s, **fpn_args) for s in scales])
+
+    def forward(self, images, need_tracker=False, tracker_mode=None, cached_trunk=None, tracker_only=False):
+        backbone_out = cached_trunk if cached_trunk is not None else self.trunk(images)
+
+        if tracker_only:
+            # Skip detector FPN when only tracker features are needed (video tracking)
+            if self.multiplex:
+                tracker_convs = self.propagation_convs if tracker_mode == "propagation" else self.interactive_convs
+            else:
+                tracker_convs = self.sam2_convs
+            tracker_features = [conv(backbone_out) for conv in tracker_convs]
+            tracker_positions = [cast_to_input(self.position_encoding(f), f) for f in tracker_features]
+            return None, None, tracker_features, tracker_positions
+
+        features = [conv(backbone_out) for conv in self.convs]
+        positions = [cast_to_input(self.position_encoding(f), f) for f in features]
+
+        if self.multiplex:
+            if tracker_mode == "propagation":
+                tracker_convs = self.propagation_convs
+            elif tracker_mode == "interactive":
+                tracker_convs = self.interactive_convs
+            else:
+                return features, positions, None, None
+        elif need_tracker:
+            tracker_convs = self.sam2_convs
+        else:
+            return features, positions, None, None
+
+        tracker_features = [conv(backbone_out) for conv in tracker_convs]
+        tracker_positions = [cast_to_input(self.position_encoding(f), f) for f in tracker_features]
+        return features, positions, tracker_features, tracker_positions

comfy/lora.py

@@ -17,6 +17,7 @@
 """
 
 from __future__ import annotations
+import comfy.memory_management
 import comfy.utils
 import comfy.model_management
 import comfy.model_base
@@ -342,6 +343,12 @@ def model_lora_keys_unet(model, key_map={}):
                 key_map["base_model.model.{}".format(key_lora)] = k  # Official base model loras
                 key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = k  # LyCORIS/LoKR format
 
+    if isinstance(model, comfy.model_base.ErnieImage):
+        for k in sdk:
+            if k.startswith("diffusion_model.") and k.endswith(".weight"):
+                key_lora = k[len("diffusion_model."):-len(".weight")]
+                key_map["transformer.{}".format(key_lora)] = k
+
     return key_map
 
 
@@ -467,3 +474,17 @@ def calculate_weight(patches, weight, key, intermediate_dtype=torch.float32, ori
             weight = old_weight
 
     return weight
+
+def prefetch_prepared_value(value, allocate_buffer, stream):
+    if isinstance(value, torch.Tensor):
+        dest = allocate_buffer(comfy.memory_management.vram_aligned_size(value))
+        comfy.model_management.cast_to_gathered([value], dest, non_blocking=True, stream=stream)
+        return comfy.memory_management.interpret_gathered_like([value], dest)[0]
+    elif isinstance(value, weight_adapter.WeightAdapterBase):
+        return type(value)(value.loaded_keys, prefetch_prepared_value(value.weights, allocate_buffer, stream))
+    elif isinstance(value, tuple):
+        return tuple(prefetch_prepared_value(item, allocate_buffer, stream) for item in value)
+    elif isinstance(value, list):
+        return [prefetch_prepared_value(item, allocate_buffer, stream) for item in value]
+
+    return value

comfy/model_base.py

@@ -52,8 +52,10 @@ import comfy.ldm.qwen_image.model
 import comfy.ldm.kandinsky5.model
 import comfy.ldm.anima.model
 import comfy.ldm.ace.ace_step15
+import comfy.ldm.cogvideo.model
 import comfy.ldm.rt_detr.rtdetr_v4
 import comfy.ldm.ernie.model
+import comfy.ldm.sam3.detector
 
 import comfy.model_management
 import comfy.patcher_extension
@@ -80,6 +82,7 @@ class ModelType(Enum):
     IMG_TO_IMG = 9
     FLOW_COSMOS = 10
     IMG_TO_IMG_FLOW = 11
+    V_PREDICTION_DDPM = 12
 
 
 def model_sampling(model_config, model_type):
@@ -114,6 +117,8 @@ def model_sampling(model_config, model_type):
         s = comfy.model_sampling.ModelSamplingCosmosRFlow
     elif model_type == ModelType.IMG_TO_IMG_FLOW:
         c = comfy.model_sampling.IMG_TO_IMG_FLOW
+    elif model_type == ModelType.V_PREDICTION_DDPM:
+        c = comfy.model_sampling.V_PREDICTION_DDPM
 
     class ModelSampling(s, c):
         pass
@@ -209,6 +214,11 @@ class BaseModel(torch.nn.Module):
         if "latent_shapes" in extra_conds:
             xc = utils.unpack_latents(xc, extra_conds.pop("latent_shapes"))
 
+        transformer_options = transformer_options.copy()
+        transformer_options["prefetch_dynamic_vbars"] = (
+            self.current_patcher is not None and self.current_patcher.is_dynamic()
+        )
+
         model_output = self.diffusion_model(xc, t, context=context, control=control, transformer_options=transformer_options, **extra_conds)
         if len(model_output) > 1 and not torch.is_tensor(model_output):
             model_output, _ = utils.pack_latents(model_output)
@@ -578,8 +588,8 @@ class Stable_Zero123(BaseModel):
     def __init__(self, model_config, model_type=ModelType.EPS, device=None, cc_projection_weight=None, cc_projection_bias=None):
         super().__init__(model_config, model_type, device=device)
         self.cc_projection = comfy.ops.manual_cast.Linear(cc_projection_weight.shape[1], cc_projection_weight.shape[0], dtype=self.get_dtype(), device=device)
-        self.cc_projection.weight.copy_(cc_projection_weight)
-        self.cc_projection.bias.copy_(cc_projection_bias)
+        self.cc_projection.weight = torch.nn.Parameter(cc_projection_weight.clone())
+        self.cc_projection.bias = torch.nn.Parameter(cc_projection_bias.clone())
 
     def extra_conds(self, **kwargs):
         out = {}
@@ -1974,3 +1984,63 @@ class ErnieImage(BaseModel):
         if cross_attn is not None:
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
         return out
+
+class SAM3(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.sam3.detector.SAM3Model)
+
+class CogVideoX(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.V_PREDICTION_DDPM, image_to_video=False, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.cogvideo.model.CogVideoXTransformer3DModel)
+        self.image_to_video = image_to_video
+
+    def concat_cond(self, **kwargs):
+        noise = kwargs.get("noise", None)
+        # Detect extra channels needed (e.g. 32 - 16 = 16 for ref latent)
+        extra_channels = self.diffusion_model.in_channels - noise.shape[1]
+        if extra_channels == 0:
+            return None
+
+        image = kwargs.get("concat_latent_image", None)
+        device = kwargs["device"]
+
+        if image is None:
+            shape = list(noise.shape)
+            shape[1] = extra_channels
+            return torch.zeros(shape, dtype=noise.dtype, layout=noise.layout, device=noise.device)
+
+        latent_dim = self.latent_format.latent_channels
+        image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
+
+        if noise.ndim == 5 and image.ndim == 5:
+            if image.shape[-3] < noise.shape[-3]:
+                image = torch.nn.functional.pad(image, (0, 0, 0, 0, 0, noise.shape[-3] - image.shape[-3]), "constant", 0)
+            elif image.shape[-3] > noise.shape[-3]:
+                image = image[:, :, :noise.shape[-3]]
+
+        for i in range(0, image.shape[1], latent_dim):
+            image[:, i:i + latent_dim] = self.process_latent_in(image[:, i:i + latent_dim])
+        image = utils.resize_to_batch_size(image, noise.shape[0])
+
+        if image.shape[1] > extra_channels:
+            image = image[:, :extra_channels]
+        elif image.shape[1] < extra_channels:
+            repeats = extra_channels // image.shape[1]
+            remainder = extra_channels % image.shape[1]
+            parts = [image] * repeats
+            if remainder > 0:
+                parts.append(image[:, :remainder])
+            image = torch.cat(parts, dim=1)
+
+        return image
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        # OFS embedding (CogVideoX 1.5 I2V), default 2.0 as used by SparkVSR
+        if self.diffusion_model.ofs_proj_dim is not None:
+            ofs = kwargs.get("ofs", None)
+            if ofs is None:
+                noise = kwargs.get("noise", None)
+                ofs = torch.full((noise.shape[0],), 2.0, device=noise.device, dtype=noise.dtype)
+            out['ofs'] = comfy.conds.CONDRegular(ofs)
+        return out

comfy/model_detection.py

@@ -490,6 +490,54 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
 
         return dit_config
 
+    if '{}blocks.0.norm1.linear.weight'.format(key_prefix) in state_dict_keys:  # CogVideoX
+        dit_config = {}
+        dit_config["image_model"] = "cogvideox"
+
+        # Extract config from weight shapes
+        norm1_weight = state_dict['{}blocks.0.norm1.linear.weight'.format(key_prefix)]
+        time_embed_dim = norm1_weight.shape[1]
+        dim = norm1_weight.shape[0] // 6
+
+        dit_config["num_attention_heads"] = dim // 64
+        dit_config["attention_head_dim"] = 64
+        dit_config["time_embed_dim"] = time_embed_dim
+        dit_config["num_layers"] = count_blocks(state_dict_keys, '{}blocks.'.format(key_prefix) + '{}.')
+
+        # Detect in_channels from patch_embed
+        patch_proj_key = '{}patch_embed.proj.weight'.format(key_prefix)
+        if patch_proj_key in state_dict_keys:
+            w = state_dict[patch_proj_key]
+            if w.ndim == 4:
+                # Conv2d: [out, in, kh, kw] — CogVideoX 1.0
+                dit_config["in_channels"] = w.shape[1]
+                dit_config["patch_size"] = w.shape[2]
+            elif w.ndim == 2:
+                # Linear: [out, in_channels * patch_size * patch_size * patch_size_t] — CogVideoX 1.5
+                dit_config["patch_size"] = 2
+                dit_config["patch_size_t"] = 2
+                dit_config["in_channels"] = w.shape[1] // (2 * 2 * 2)  # 256 // 8 = 32
+
+        text_proj_key = '{}patch_embed.text_proj.weight'.format(key_prefix)
+        if text_proj_key in state_dict_keys:
+            dit_config["text_embed_dim"] = state_dict[text_proj_key].shape[1]
+
+        # Detect OFS embedding
+        ofs_key = '{}ofs_embedding_linear_1.weight'.format(key_prefix)
+        if ofs_key in state_dict_keys:
+            dit_config["ofs_embed_dim"] = state_dict[ofs_key].shape[1]
+
+        # Detect positional embedding type
+        pos_key = '{}patch_embed.pos_embedding'.format(key_prefix)
+        if pos_key in state_dict_keys:
+            dit_config["use_learned_positional_embeddings"] = True
+            dit_config["use_rotary_positional_embeddings"] = False
+        else:
+            dit_config["use_learned_positional_embeddings"] = False
+            dit_config["use_rotary_positional_embeddings"] = True
+
+        return dit_config
+
     if '{}head.modulation'.format(key_prefix) in state_dict_keys:  # Wan 2.1
         dit_config = {}
         dit_config["image_model"] = "wan2.1"
@@ -718,6 +766,14 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config["image_model"] = "ernie"
         return dit_config
 
+    if 'detector.backbone.vision_backbone.trunk.blocks.0.attn.qkv.weight' in state_dict_keys: # SAM3 / SAM3.1
+        if 'detector.transformer.decoder.query_embed.weight' in state_dict_keys:
+            dit_config = {}
+            dit_config["image_model"] = "SAM3"
+            if 'detector.backbone.vision_backbone.propagation_convs.0.conv_1x1.weight' in state_dict_keys:
+                dit_config["image_model"] = "SAM31"
+            return dit_config
+
     if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:
         return None
 
@@ -873,6 +929,10 @@ def model_config_from_unet(state_dict, unet_key_prefix, use_base_if_no_match=Fal
     return model_config
 
 def unet_prefix_from_state_dict(state_dict):
+    # SAM3: detector.* and tracker.* at top level, no common prefix
+    if any(k.startswith("detector.") for k in state_dict) and any(k.startswith("tracker.") for k in state_dict):
+        return ""
+
     candidates = ["model.diffusion_model.", #ldm/sgm models
                   "model.model.", #audio models
                   "net.", #cosmos

comfy/model_management.py

@@ -31,6 +31,7 @@ from contextlib import nullcontext
 import comfy.memory_management
 import comfy.utils
 import comfy.quant_ops
+import comfy_aimdo.vram_buffer
 
 class VRAMState(Enum):
     DISABLED = 0    #No vram present: no need to move models to vram
@@ -112,10 +113,6 @@ if args.directml is not None:
     # torch_directml.disable_tiled_resources(True)
     lowvram_available = False #TODO: need to find a way to get free memory in directml before this can be enabled by default.
 
-try:
-    import intel_extension_for_pytorch as ipex  # noqa: F401
-except:
-    pass
 
 try:
     _ = torch.xpu.device_count()
@@ -583,9 +580,6 @@ class LoadedModel:
 
         real_model = self.model.model
 
-        if is_intel_xpu() and not args.disable_ipex_optimize and 'ipex' in globals() and real_model is not None:
-            with torch.no_grad():
-                real_model = ipex.optimize(real_model.eval(), inplace=True, graph_mode=True, concat_linear=True)
 
         self.real_model = weakref.ref(real_model)
         self.model_finalizer = weakref.finalize(real_model, cleanup_models)
@@ -663,6 +657,7 @@ def minimum_inference_memory():
 
 def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, pins_required=0, ram_required=0):
     cleanup_models_gc()
+    comfy.memory_management.extra_ram_release(max(pins_required, ram_required))
     unloaded_model = []
     can_unload = []
     unloaded_models = []
@@ -1181,6 +1176,10 @@ stream_counters = {}
 
 STREAM_CAST_BUFFERS = {}
 LARGEST_CASTED_WEIGHT = (None, 0)
+STREAM_AIMDO_CAST_BUFFERS = {}
+LARGEST_AIMDO_CASTED_WEIGHT = (None, 0)
+
+DEFAULT_AIMDO_CAST_BUFFER_RESERVATION_SIZE = 16 * 1024 ** 3
 
 def get_cast_buffer(offload_stream, device, size, ref):
     global LARGEST_CASTED_WEIGHT
@@ -1214,13 +1213,26 @@ def get_cast_buffer(offload_stream, device, size, ref):
 
     return cast_buffer
 
+def get_aimdo_cast_buffer(offload_stream, device):
+    cast_buffer = STREAM_AIMDO_CAST_BUFFERS.get(offload_stream, None)
+    if cast_buffer is None:
+        cast_buffer = comfy_aimdo.vram_buffer.VRAMBuffer(DEFAULT_AIMDO_CAST_BUFFER_RESERVATION_SIZE, device.index)
+        STREAM_AIMDO_CAST_BUFFERS[offload_stream] = cast_buffer
+
+    return cast_buffer
 def reset_cast_buffers():
     global LARGEST_CASTED_WEIGHT
+    global LARGEST_AIMDO_CASTED_WEIGHT
+
     LARGEST_CASTED_WEIGHT = (None, 0)
-    for offload_stream in STREAM_CAST_BUFFERS:
-        offload_stream.synchronize()
+    LARGEST_AIMDO_CASTED_WEIGHT = (None, 0)
+    for offload_stream in set(STREAM_CAST_BUFFERS) | set(STREAM_AIMDO_CAST_BUFFERS):
+        if offload_stream is not None:
+            offload_stream.synchronize()
     synchronize()
+
     STREAM_CAST_BUFFERS.clear()
+    STREAM_AIMDO_CAST_BUFFERS.clear()
     soft_empty_cache()
 
 def get_offload_stream(device):
@@ -1580,10 +1592,7 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True, ma
         return False
 
     if is_intel_xpu():
-        if torch_version_numeric < (2, 3):
-            return True
-        else:
-            return torch.xpu.get_device_properties(device).has_fp16
+        return torch.xpu.get_device_properties(device).has_fp16
 
     if is_ascend_npu():
         return True
@@ -1649,10 +1658,7 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma
         return False
 
     if is_intel_xpu():
-        if torch_version_numeric < (2, 3):
-            return True
-        else:
-            return torch.xpu.is_bf16_supported()
+        return torch.xpu.is_bf16_supported()
 
     if is_ascend_npu():
         return True
@@ -1783,6 +1789,7 @@ def soft_empty_cache(force=False):
     if cpu_state == CPUState.MPS:
         torch.mps.empty_cache()
     elif is_intel_xpu():
+        torch.xpu.synchronize()
         torch.xpu.empty_cache()
     elif is_ascend_npu():
         torch.npu.empty_cache()
@@ -1801,7 +1808,7 @@ def debug_memory_summary():
         return torch.cuda.memory.memory_summary()
     return ""
 
-class InterruptProcessingException(Exception):
+class InterruptProcessingException(BaseException):
     pass
 
 interrupt_processing_mutex = threading.RLock()

comfy/model_patcher.py

@@ -31,6 +31,7 @@ import comfy.float
 import comfy.hooks
 import comfy.lora
 import comfy.model_management
+import comfy.ops
 import comfy.patcher_extension
 import comfy.utils
 from comfy.comfy_types import UnetWrapperFunction
@@ -120,9 +121,20 @@ class LowVramPatch:
         self.patches = patches
         self.convert_func = convert_func # TODO: remove
         self.set_func = set_func
+        self.prepared_patches = None
+
+    def prepare(self, allocate_buffer, stream):
+        self.prepared_patches = [
+            (patch[0], comfy.lora.prefetch_prepared_value(patch[1], allocate_buffer, stream), patch[2], patch[3], patch[4])
+            for patch in self.patches[self.key]
+        ]
+
+    def clear_prepared(self):
+        self.prepared_patches = None
 
     def __call__(self, weight):
-        return comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=weight.dtype)
+        patches = self.prepared_patches if self.prepared_patches is not None else self.patches[self.key]
+        return comfy.lora.calculate_weight(patches, weight, self.key, intermediate_dtype=weight.dtype)
 
 LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR = 2
 
@@ -685,9 +697,9 @@ class ModelPatcher:
                         sd.pop(k)
             return sd
 
-    def patch_weight_to_device(self, key, device_to=None, inplace_update=False, return_weight=False):
+    def patch_weight_to_device(self, key, device_to=None, inplace_update=False, return_weight=False, force_cast=False):
         weight, set_func, convert_func = get_key_weight(self.model, key)
-        if key not in self.patches:
+        if key not in self.patches and not force_cast:
             return weight
 
         inplace_update = self.weight_inplace_update or inplace_update
@@ -695,7 +707,7 @@ class ModelPatcher:
         if key not in self.backup and not return_weight:
             self.backup[key] = collections.namedtuple('Dimension', ['weight', 'inplace_update'])(weight.to(device=self.offload_device, copy=inplace_update), inplace_update)
 
-        temp_dtype = comfy.model_management.lora_compute_dtype(device_to)
+        temp_dtype = comfy.model_management.lora_compute_dtype(device_to) if key in self.patches else None
         if device_to is not None:
             temp_weight = comfy.model_management.cast_to_device(weight, device_to, temp_dtype, copy=True)
         else:
@@ -703,9 +715,10 @@ class ModelPatcher:
         if convert_func is not None:
             temp_weight = convert_func(temp_weight, inplace=True)
 
-        out_weight = comfy.lora.calculate_weight(self.patches[key], temp_weight, key)
+        out_weight = comfy.lora.calculate_weight(self.patches[key], temp_weight, key) if key in self.patches else temp_weight
         if set_func is None:
-            out_weight = comfy.float.stochastic_rounding(out_weight, weight.dtype, seed=comfy.utils.string_to_seed(key))
+            if key in self.patches:
+                out_weight = comfy.float.stochastic_rounding(out_weight, weight.dtype, seed=comfy.utils.string_to_seed(key))
             if return_weight:
                 return out_weight
             elif inplace_update:
@@ -855,7 +868,9 @@ class ModelPatcher:
                     if m.comfy_patched_weights == True:
                         continue
 
-                for param in params:
+                for param, param_value in params.items():
+                    if hasattr(m, "comfy_cast_weights") and getattr(param_value, "is_meta", False):
+                        comfy.ops.disable_weight_init._zero_init_parameter(m, param)
                     key = key_param_name_to_key(n, param)
                     self.unpin_weight(key)
                     self.patch_weight_to_device(key, device_to=device_to)
@@ -1584,7 +1599,7 @@ class ModelPatcherDynamic(ModelPatcher):
                     key = key_param_name_to_key(n, param_key)
                     if key in self.backup:
                         comfy.utils.set_attr_param(self.model, key, self.backup[key].weight)
-                    self.patch_weight_to_device(key, device_to=device_to)
+                    self.patch_weight_to_device(key, device_to=device_to, force_cast=True)
                     weight, _, _ = get_key_weight(self.model, key)
                     if weight is not None:
                         self.model.model_loaded_weight_memory += weight.numel() * weight.element_size()
@@ -1609,6 +1624,10 @@ class ModelPatcherDynamic(ModelPatcher):
                             m._v = vbar.alloc(v_weight_size)
                         allocated_size += v_weight_size
 
+                    for param in params:
+                        if param not in ("weight", "bias"):
+                            force_load_param(self, param, device_to)
+
                 else:
                     for param in params:
                         key = key_param_name_to_key(n, param)

comfy/model_prefetch.py

@@ -0,0 +1,65 @@
+import comfy_aimdo.model_vbar
+import comfy.model_management
+import comfy.ops
+
+PREFETCH_QUEUES = []
+
+def cleanup_prefetched_modules(comfy_modules):
+    for s in comfy_modules:
+        prefetch = getattr(s, "_prefetch", None)
+        if prefetch is None:
+            continue
+        for param_key in ("weight", "bias"):
+            lowvram_fn = getattr(s, param_key + "_lowvram_function", None)
+            if lowvram_fn is not None:
+                lowvram_fn.clear_prepared()
+        if prefetch["signature"] is not None:
+            comfy_aimdo.model_vbar.vbar_unpin(s._v)
+        delattr(s, "_prefetch")
+
+def cleanup_prefetch_queues():
+    global PREFETCH_QUEUES
+
+    for queue in PREFETCH_QUEUES:
+        for entry in queue:
+            if entry is None or not isinstance(entry, tuple):
+                continue
+            _, prefetch_state = entry
+            comfy_modules = prefetch_state[1]
+            if comfy_modules is not None:
+                cleanup_prefetched_modules(comfy_modules)
+    PREFETCH_QUEUES = []
+
+def prefetch_queue_pop(queue, device, module):
+    if queue is None:
+        return
+
+    consumed = queue.pop(0)
+    if consumed is not None:
+        offload_stream, prefetch_state = consumed
+        offload_stream.wait_stream(comfy.model_management.current_stream(device))
+        _, comfy_modules = prefetch_state
+        if comfy_modules is not None:
+            cleanup_prefetched_modules(comfy_modules)
+
+    prefetch = queue[0]
+    if prefetch is not None:
+        comfy_modules = []
+        for s in prefetch.modules():
+            if hasattr(s, "_v"):
+                comfy_modules.append(s)
+
+        offload_stream = comfy.ops.cast_modules_with_vbar(comfy_modules, None, device, None, True)
+        comfy.model_management.sync_stream(device, offload_stream)
+        queue[0] = (offload_stream, (prefetch, comfy_modules))
+
+def make_prefetch_queue(queue, device, transformer_options):
+    if (not transformer_options.get("prefetch_dynamic_vbars", False)
+        or comfy.model_management.NUM_STREAMS == 0
+        or comfy.model_management.is_device_cpu(device)
+        or not comfy.model_management.device_supports_non_blocking(device)):
+        return None
+
+    queue = [None] + queue + [None]
+    PREFETCH_QUEUES.append(queue)
+    return queue

comfy/model_sampling.py

@@ -54,6 +54,30 @@ class V_PREDICTION(EPS):
         sigma = reshape_sigma(sigma, model_output.ndim)
         return model_input * self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) - model_output * sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2) ** 0.5
 
+class V_PREDICTION_DDPM:
+    """CogVideoX v-prediction: model receives raw x_t (unscaled), predicts velocity v.
+    x_0 = sqrt(alpha) * x_t - sqrt(1-alpha) * v
+        = x_t / sqrt(sigma^2 + 1) - v * sigma / sqrt(sigma^2 + 1)
+    """
+    def calculate_input(self, sigma, noise):
+        return noise
+
+    def calculate_denoised(self, sigma, model_output, model_input):
+        sigma = reshape_sigma(sigma, model_output.ndim)
+        return model_input / (sigma ** 2 + 1.0) ** 0.5 - model_output * sigma / (sigma ** 2 + 1.0) ** 0.5
+
+    def noise_scaling(self, sigma, noise, latent_image, max_denoise=False):
+        sigma = reshape_sigma(sigma, noise.ndim)
+        if max_denoise:
+            noise = noise * torch.sqrt(1.0 + sigma ** 2.0)
+        else:
+            noise = noise * sigma
+        noise += latent_image
+        return noise
+
+    def inverse_noise_scaling(self, sigma, latent):
+        return latent
+
 class EDM(V_PREDICTION):
     def calculate_denoised(self, sigma, model_output, model_input):
         sigma = reshape_sigma(sigma, model_output.ndim)

comfy/ops.py

@@ -79,37 +79,68 @@ def cast_to_input(weight, input, non_blocking=False, copy=True):
     return comfy.model_management.cast_to(weight, input.dtype, input.device, non_blocking=non_blocking, copy=copy)
 
 
-def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compute_dtype, want_requant):
+def materialize_meta_param(s, param_keys):
+    for param_key in param_keys:
+        param = getattr(s, param_key, None)
+        if param is not None and getattr(param, "is_meta", False):
+            setattr(s, param_key, torch.nn.Parameter(torch.zeros(param.shape, dtype=param.dtype), requires_grad=param.requires_grad))
 
-    #vbar doesn't support CPU weights, but some custom nodes have weird paths
-    #that might switch the layer to the CPU and expect it to work. We have to take
-    #a clone conservatively as we are mmapped and some SFT files are packed misaligned
-    #If you are a custom node author reading this, please move your layer to the GPU
-    #or declare your ModelPatcher as CPU in the first place.
-    if comfy.model_management.is_device_cpu(device):
-        weight = s.weight.to(dtype=dtype, copy=True)
-        if isinstance(weight, QuantizedTensor):
-            weight = weight.dequantize()
-        bias = None
-        if s.bias is not None:
-            bias = s.bias.to(dtype=bias_dtype, copy=True)
-        return weight, bias, (None, None, None)
 
+# FIXME: add n=1 cache hit fast path
+def cast_modules_with_vbar(comfy_modules, dtype, device, bias_dtype, non_blocking):
     offload_stream = None
-    xfer_dest = None
+    cast_buffer = None
+    cast_buffer_offset = 0
+
+    def ensure_offload_stream(module, required_size, check_largest):
+        nonlocal offload_stream
+        nonlocal cast_buffer
+
+        if offload_stream is None:
+            offload_stream = comfy.model_management.get_offload_stream(device)
+        if offload_stream is None or not check_largest or len(comfy_modules) != 1:
+            return
+
+        current_size = 0 if cast_buffer is None else cast_buffer.size()
+        if current_size < required_size and module is comfy.model_management.LARGEST_AIMDO_CASTED_WEIGHT[0]:
+            offload_stream = comfy.model_management.get_offload_stream(device)
+            cast_buffer = None
+        if required_size > comfy.model_management.LARGEST_AIMDO_CASTED_WEIGHT[1]:
+            comfy.model_management.LARGEST_AIMDO_CASTED_WEIGHT = (module, required_size)
+
+    def get_cast_buffer(buffer_size):
+        nonlocal offload_stream
+        nonlocal cast_buffer
+        nonlocal cast_buffer_offset
+
+        if buffer_size == 0:
+            return None
+
+        if offload_stream is None:
+            return torch.empty((buffer_size,), dtype=torch.uint8, device=device)
+
+        cast_buffer = comfy.model_management.get_aimdo_cast_buffer(offload_stream, device)
+        buffer = comfy_aimdo.torch.aimdo_to_tensor(cast_buffer.get(buffer_size, cast_buffer_offset), device)
+        cast_buffer_offset += buffer_size
+        return buffer
+
+    for s in comfy_modules:
+        signature = comfy_aimdo.model_vbar.vbar_fault(s._v)
+        resident = comfy_aimdo.model_vbar.vbar_signature_compare(signature, s._v_signature)
+        prefetch = {
+            "signature": signature,
+            "resident": resident,
+        }
 
-    signature = comfy_aimdo.model_vbar.vbar_fault(s._v)
-    resident = comfy_aimdo.model_vbar.vbar_signature_compare(signature, s._v_signature)
-    if signature is not None:
         if resident:
-            weight = s._v_weight
-            bias = s._v_bias
-        else:
-            xfer_dest = comfy_aimdo.torch.aimdo_to_tensor(s._v, device)
+            s._prefetch = prefetch
+            continue
 
-    if not resident:
+        materialize_meta_param(s, ["weight", "bias"])
+        xfer_dest = comfy_aimdo.torch.aimdo_to_tensor(s._v, device) if signature is not None else None
         cast_geometry = comfy.memory_management.tensors_to_geometries([ s.weight, s.bias ])
         cast_dest = None
+        needs_cast = False
 
         xfer_source = [ s.weight, s.bias ]
 
@@ -121,22 +152,15 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
             if data is None:
                 continue
             if data.dtype != geometry.dtype:
+                needs_cast = True
                 cast_dest = xfer_dest
-                if cast_dest is None:
-                    cast_dest = torch.empty((comfy.memory_management.vram_aligned_size(cast_geometry),), dtype=torch.uint8, device=device)
                 xfer_dest = None
                 break
 
         dest_size = comfy.memory_management.vram_aligned_size(xfer_source)
-        offload_stream = comfy.model_management.get_offload_stream(device)
-        if xfer_dest is None and offload_stream is not None:
-                xfer_dest = comfy.model_management.get_cast_buffer(offload_stream, device, dest_size, s)
-                if xfer_dest is None:
-                    offload_stream = comfy.model_management.get_offload_stream(device)
-                    xfer_dest = comfy.model_management.get_cast_buffer(offload_stream, device, dest_size, s)
+        ensure_offload_stream(s, dest_size if xfer_dest is None else 0, True)
         if xfer_dest is None:
-            xfer_dest = torch.empty((dest_size,), dtype=torch.uint8, device=device)
-            offload_stream = None
+            xfer_dest = get_cast_buffer(dest_size)
 
         if signature is None and pin is None:
             comfy.pinned_memory.pin_memory(s)
@@ -149,27 +173,54 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
             xfer_source = [ pin ]
         #send it over
         comfy.model_management.cast_to_gathered(xfer_source, xfer_dest, non_blocking=non_blocking, stream=offload_stream)
-        comfy.model_management.sync_stream(device, offload_stream)
 
-        if cast_dest is not None:
+        for param_key in ("weight", "bias"):
+            lowvram_fn = getattr(s, param_key + "_lowvram_function", None)
+            if lowvram_fn is not None:
+                ensure_offload_stream(s, cast_buffer_offset, False)
+                lowvram_fn.prepare(lambda size: get_cast_buffer(size), offload_stream)
+
+        prefetch["xfer_dest"] = xfer_dest
+        prefetch["cast_dest"] = cast_dest
+        prefetch["cast_geometry"] = cast_geometry
+        prefetch["needs_cast"] = needs_cast
+        s._prefetch = prefetch
+
+    return offload_stream
+
+
+def resolve_cast_module_with_vbar(s, dtype, device, bias_dtype, compute_dtype, want_requant):
+
+    prefetch = getattr(s, "_prefetch", None)
+
+    if prefetch["resident"]:
+        weight = s._v_weight
+        bias = s._v_bias
+    else:
+        xfer_dest = prefetch["xfer_dest"]
+        if prefetch["needs_cast"]:
+            cast_dest = prefetch["cast_dest"] if prefetch["cast_dest"] is not None else torch.empty((comfy.memory_management.vram_aligned_size(prefetch["cast_geometry"]),), dtype=torch.uint8, device=device)
             for pre_cast, post_cast in zip(comfy.memory_management.interpret_gathered_like([s.weight, s.bias ], xfer_dest),
-                                           comfy.memory_management.interpret_gathered_like(cast_geometry, cast_dest)):
+                                           comfy.memory_management.interpret_gathered_like(prefetch["cast_geometry"], cast_dest)):
                 if post_cast is not None:
                     post_cast.copy_(pre_cast)
             xfer_dest = cast_dest
 
-        params = comfy.memory_management.interpret_gathered_like(cast_geometry, xfer_dest)
+        params = comfy.memory_management.interpret_gathered_like(prefetch["cast_geometry"], xfer_dest)
         weight = params[0]
         bias = params[1]
-        if signature is not None:
+        if prefetch["signature"] is not None:
             s._v_weight = weight
             s._v_bias = bias
-        s._v_signature=signature
+        s._v_signature = prefetch["signature"]
 
     def post_cast(s, param_key, x, dtype, resident, update_weight):
         lowvram_fn = getattr(s, param_key + "_lowvram_function", None)
         fns = getattr(s, param_key + "_function", [])
 
+        if x is None:
+            return None
+
         orig = x
 
         def to_dequant(tensor, dtype):
@@ -197,14 +248,12 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
             x = f(x)
         return x
 
-    update_weight = signature is not None
+    update_weight = prefetch["signature"] is not None
+    weight = post_cast(s, "weight", weight, dtype, prefetch["resident"], update_weight)
+    if bias is not None:
+        bias = post_cast(s, "bias", bias, bias_dtype, prefetch["resident"], update_weight)
 
-    weight = post_cast(s, "weight", weight, dtype, resident, update_weight)
-    if s.bias is not None:
-        bias = post_cast(s, "bias", bias, bias_dtype, resident, update_weight)
-
-    #FIXME: weird offload return protocol
-    return weight, bias, (offload_stream, device if signature is not None else None, None)
+    return weight, bias
 
 
 def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, offloadable=False, compute_dtype=None, want_requant=False):
@@ -222,10 +271,46 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of
         if device is None:
             device = input.device
 
+    def format_return(result, offloadable):
+        weight, bias, offload_stream = result
+        return (weight, bias, offload_stream) if offloadable else (weight, bias)
+
     non_blocking = comfy.model_management.device_supports_non_blocking(device)
 
     if hasattr(s, "_v"):
-        return cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compute_dtype, want_requant)
+
+        #vbar doesn't support CPU weights, but some custom nodes have weird paths
+        #that might switch the layer to the CPU and expect it to work. We have to take
+        #a clone conservatively as we are mmapped and some SFT files are packed misaligned
+        #If you are a custom node author reading this, please move your layer to the GPU
+        #or declare your ModelPatcher as CPU in the first place.
+        if comfy.model_management.is_device_cpu(device):
+            materialize_meta_param(s, ["weight", "bias"])
+            weight = s.weight.to(dtype=dtype, copy=True)
+            if isinstance(weight, QuantizedTensor):
+                weight = weight.dequantize()
+            bias = s.bias.to(dtype=bias_dtype, copy=True) if s.bias is not None else None
+            return format_return((weight, bias, (None, None, None)), offloadable)
+
+        prefetched = hasattr(s, "_prefetch")
+        offload_stream = None
+        offload_device = None
+        if not prefetched:
+            offload_stream = cast_modules_with_vbar([s], dtype, device, bias_dtype, non_blocking)
+            comfy.model_management.sync_stream(device, offload_stream)
+
+        weight, bias = resolve_cast_module_with_vbar(s, dtype, device, bias_dtype, compute_dtype, want_requant)
+
+        if not prefetched:
+            if getattr(s, "_prefetch")["signature"] is not None:
+                offload_device = device
+            for param_key in ("weight", "bias"):
+                lowvram_fn = getattr(s, param_key + "_lowvram_function", None)
+                if lowvram_fn is not None:
+                    lowvram_fn.clear_prepared()
+            delattr(s, "_prefetch")
+        return format_return((weight, bias, (offload_stream, offload_device, None)), offloadable)
+
 
     if offloadable and (device != s.weight.device or
                         (s.bias is not None and device != s.bias.device)):
@@ -272,11 +357,7 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of
         for f in s.weight_function:
             weight = f(weight)
 
-    if offloadable:
-        return weight, bias, (offload_stream, weight_a, bias_a)
-    else:
-        #Legacy function signature
-        return weight, bias
+    return format_return((weight, bias, (offload_stream, weight_a, bias_a)), offloadable)
 
 
 def uncast_bias_weight(s, weight, bias, offload_stream):
@@ -306,6 +387,12 @@ class CastWeightBiasOp:
     bias_function = []
 
 class disable_weight_init:
+    @staticmethod
+    def _zero_init_parameter(module, name):
+        param = getattr(module, name)
+        device = None if getattr(param, "is_meta", False) else param.device
+        setattr(module, name, torch.nn.Parameter(torch.zeros(param.shape, device=device, dtype=param.dtype), requires_grad=False))
+
     @staticmethod
     def _lazy_load_from_state_dict(module, state_dict, prefix, local_metadata,
                                    missing_keys, unexpected_keys, weight_shape,
@@ -1159,6 +1246,93 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                         self._buffers[key] = fn(buf)
                 return self
 
+        class Embedding(manual_cast.Embedding):
+            def _load_from_state_dict(self, state_dict, prefix, local_metadata,
+                                    strict, missing_keys, unexpected_keys, error_msgs):
+                weight_key = f"{prefix}weight"
+                layer_conf = state_dict.pop(f"{prefix}comfy_quant", None)
+                if layer_conf is not None:
+                    layer_conf = json.loads(layer_conf.numpy().tobytes())
+
+                # Only fp8 makes sense for embeddings (per-row dequant via index select).
+                # Block-scaled formats (NVFP4, MXFP8) can't do per-row lookup efficiently.
+                quant_format = layer_conf.get("format", None) if layer_conf is not None else None
+                if quant_format in ["float8_e4m3fn", "float8_e5m2"] and weight_key in state_dict:
+                    self.quant_format = quant_format
+                    qconfig = QUANT_ALGOS[quant_format]
+                    layout_cls = get_layout_class(qconfig["comfy_tensor_layout"])
+                    weight = state_dict.pop(weight_key)
+                    manually_loaded_keys = [weight_key]
+
+                    scale_key = f"{prefix}weight_scale"
+                    scale = state_dict.pop(scale_key, None)
+                    if scale is not None:
+                        scale = scale.float()
+                        manually_loaded_keys.append(scale_key)
+
+                    params = layout_cls.Params(
+                        scale=scale if scale is not None else torch.ones((), dtype=torch.float32),
+                        orig_dtype=MixedPrecisionOps._compute_dtype,
+                        orig_shape=(self.num_embeddings, self.embedding_dim),
+                    )
+                    self.weight = torch.nn.Parameter(
+                        QuantizedTensor(weight.to(dtype=qconfig["storage_t"]), qconfig["comfy_tensor_layout"], params),
+                        requires_grad=False)
+
+                    super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs)
+                    for k in manually_loaded_keys:
+                        if k in missing_keys:
+                            missing_keys.remove(k)
+                else:
+                    if layer_conf is not None:
+                        state_dict[f"{prefix}comfy_quant"] = torch.tensor(list(json.dumps(layer_conf).encode('utf-8')), dtype=torch.uint8)
+                    super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs)
+
+            def state_dict(self, *args, destination=None, prefix="", **kwargs):
+                if destination is not None:
+                    sd = destination
+                else:
+                    sd = {}
+
+                if not hasattr(self, 'weight') or self.weight is None:
+                    return sd
+
+                if isinstance(self.weight, QuantizedTensor):
+                    sd_out = self.weight.state_dict("{}weight".format(prefix))
+                    for k in sd_out:
+                        sd[k] = sd_out[k]
+
+                    quant_conf = {"format": self.quant_format}
+                    sd["{}comfy_quant".format(prefix)] = torch.tensor(list(json.dumps(quant_conf).encode('utf-8')), dtype=torch.uint8)
+                else:
+                    sd["{}weight".format(prefix)] = self.weight
+                return sd
+
+            def forward_comfy_cast_weights(self, input, out_dtype=None):
+                weight = self.weight
+
+                # Optimized path: lookup in fp8, dequantize only the selected rows.
+                if isinstance(weight, QuantizedTensor) and len(self.weight_function) == 0:
+                    qdata, _, offload_stream = cast_bias_weight(self, device=input.device, dtype=weight.dtype, offloadable=True)
+                    if isinstance(qdata, QuantizedTensor):
+                        scale = qdata._params.scale
+                        qdata = qdata._qdata
+                    else:
+                        scale = None
+
+                    x = torch.nn.functional.embedding(
+                        input, qdata, self.padding_idx, self.max_norm,
+                        self.norm_type, self.scale_grad_by_freq, self.sparse)
+                    uncast_bias_weight(self, qdata, None, offload_stream)
+                    target_dtype = out_dtype if out_dtype is not None else weight._params.orig_dtype
+                    x = x.to(dtype=target_dtype)
+                    if scale is not None and scale != 1.0:
+                        x = x * scale.to(dtype=target_dtype)
+                    return x
+
+                # Fallback for non-quantized or weight_function (LoRA) case
+                return super().forward_comfy_cast_weights(input, out_dtype=out_dtype)
+
     return MixedPrecisionOps
 
 def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, model_config=None):

comfy/pinned_memory.py

@@ -2,7 +2,6 @@ import comfy.model_management
 import comfy.memory_management
 import comfy_aimdo.host_buffer
 import comfy_aimdo.torch
-import psutil
 
 from comfy.cli_args import args
 
@@ -12,11 +11,6 @@ def get_pin(module):
 def pin_memory(module):
     if module.pin_failed or args.disable_pinned_memory or get_pin(module) is not None:
         return
-    #FIXME: This is a RAM cache trigger event
-    ram_headroom = comfy.memory_management.RAM_CACHE_HEADROOM
-    #we split the difference and assume half the RAM cache headroom is for us
-    if ram_headroom > 0 and psutil.virtual_memory().available < (ram_headroom * 0.5):
-        comfy.memory_management.extra_ram_release(ram_headroom)
 
     size = comfy.memory_management.vram_aligned_size([ module.weight, module.bias ])
 

comfy/quant_ops.py

@@ -1,6 +1,8 @@
 import torch
 import logging
 
+from comfy.cli_args import args
+
 try:
     import comfy_kitchen as ck
     from comfy_kitchen.tensor import (
@@ -21,7 +23,15 @@ try:
             ck.registry.disable("cuda")
             logging.warning("WARNING: You need pytorch with cu130 or higher to use optimized CUDA operations.")
 
-    ck.registry.disable("triton")
+    if args.enable_triton_backend:
+        try:
+            import triton
+            logging.info("Found triton %s. Enabling comfy-kitchen triton backend.", triton.__version__)
+        except ImportError as e:
+            logging.error(f"Failed to import triton, Error: {e}, the comfy-kitchen triton backend will not be available.")
+            ck.registry.disable("triton")
+    else:
+        ck.registry.disable("triton")
     for k, v in ck.list_backends().items():
         logging.info(f"Found comfy_kitchen backend {k}: {v}")
 except ImportError as e:

comfy/rmsnorm.py

@@ -3,6 +3,7 @@ import comfy.model_management
 
 RMSNorm = torch.nn.RMSNorm
 
+# Note: torch's fused F.rms_norm is faster but produces slightly different output than manual implementations (rsqrt/reduction rounding).
 def rms_norm(x, weight=None, eps=1e-6):
     if weight is None:
         return torch.nn.functional.rms_norm(x, (x.shape[-1],), eps=eps)

comfy/sd.py

@@ -18,6 +18,7 @@ import comfy.ldm.wan.vae
 import comfy.ldm.wan.vae2_2
 import comfy.ldm.hunyuan3d.vae
 import comfy.ldm.ace.vae.music_dcae_pipeline
+import comfy.ldm.cogvideo.vae
 import comfy.ldm.hunyuan_video.vae
 import comfy.ldm.mmaudio.vae.autoencoder
 import comfy.pixel_space_convert
@@ -64,6 +65,7 @@ import comfy.text_encoders.ace15
 import comfy.text_encoders.longcat_image
 import comfy.text_encoders.qwen35
 import comfy.text_encoders.ernie
+import comfy.text_encoders.gemma4
 
 import comfy.model_patcher
 import comfy.lora
@@ -478,7 +480,10 @@ class VAE:
                                                             encoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Encoder", 'params': encoder_config},
                                                             decoder_config={'target': "comfy.ldm.modules.temporal_ae.VideoDecoder", 'params': decoder_config})
             elif "taesd_decoder.1.weight" in sd:
-                self.latent_channels = sd["taesd_decoder.1.weight"].shape[1]
+                if isinstance(metadata, dict) and "tae_latent_channels" in metadata:
+                    self.latent_channels = metadata["tae_latent_channels"]
+                else:
+                    self.latent_channels = sd["taesd_decoder.1.weight"].shape[1]
                 self.first_stage_model = comfy.taesd.taesd.TAESD(latent_channels=self.latent_channels)
             elif "vquantizer.codebook.weight" in sd: #VQGan: stage a of stable cascade
                 self.first_stage_model = StageA()
@@ -652,6 +657,17 @@ class VAE:
 
                 self.memory_used_encode = lambda shape, dtype: (1400 * 9 * shape[-2] * shape[-1]) * model_management.dtype_size(dtype)
                 self.memory_used_decode = lambda shape, dtype: (3600 * 4 * shape[-2] * shape[-1] * 16 * 16) * model_management.dtype_size(dtype)
+            elif "decoder.conv_in.conv.weight" in sd and "decoder.mid_block.resnets.0.norm1.norm_layer.weight" in sd:  # CogVideoX VAE
+                self.upscale_ratio = (lambda a: max(0, a * 4 - 3), 8, 8)
+                self.upscale_index_formula = (4, 8, 8)
+                self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 8, 8)
+                self.downscale_index_formula = (4, 8, 8)
+                self.latent_dim = 3
+                self.latent_channels = sd["encoder.conv_out.conv.weight"].shape[0] // 2
+                self.first_stage_model = comfy.ldm.cogvideo.vae.AutoencoderKLCogVideoX(latent_channels=self.latent_channels)
+                self.memory_used_decode = lambda shape, dtype: (2800 * max(2, ((shape[2] - 1) * 4) + 1) * shape[3] * shape[4] * (8 * 8)) * model_management.dtype_size(dtype)
+                self.memory_used_encode = lambda shape, dtype: (1400 * max(1, shape[2]) * shape[3] * shape[4]) * model_management.dtype_size(dtype)
+                self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32]
             elif "decoder.conv_in.conv.weight" in sd:
                 ddconfig = {'double_z': True, 'z_channels': 4, 'resolution': 256, 'in_channels': 3, 'out_ch': 3, 'ch': 128, 'ch_mult': [1, 2, 4, 4], 'num_res_blocks': 2, 'attn_resolutions': [], 'dropout': 0.0}
                 ddconfig["conv3d"] = True
@@ -1256,6 +1272,9 @@ class TEModel(Enum):
     QWEN35_9B = 26
     QWEN35_27B = 27
     MINISTRAL_3_3B = 28
+    GEMMA_4_E4B = 29
+    GEMMA_4_E2B = 30
+    GEMMA_4_31B = 31
 
 
 def detect_te_model(sd):
@@ -1281,6 +1300,12 @@ def detect_te_model(sd):
             return TEModel.BYT5_SMALL_GLYPH
         return TEModel.T5_BASE
     if 'model.layers.0.post_feedforward_layernorm.weight' in sd:
+        if 'model.layers.59.self_attn.q_norm.weight' in sd:
+            return TEModel.GEMMA_4_31B
+        if 'model.layers.41.self_attn.q_norm.weight' in sd and 'model.layers.47.self_attn.q_norm.weight' not in sd:
+            return TEModel.GEMMA_4_E4B
+        if 'model.layers.34.self_attn.q_norm.weight' in sd and 'model.layers.41.self_attn.q_norm.weight' not in sd:
+            return TEModel.GEMMA_4_E2B
         if 'model.layers.47.self_attn.q_norm.weight' in sd:
             return TEModel.GEMMA_3_12B
         if 'model.layers.0.self_attn.q_norm.weight' in sd:
@@ -1420,6 +1445,13 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
             else:
                 clip_target.clip = comfy.text_encoders.sa_t5.SAT5Model
                 clip_target.tokenizer = comfy.text_encoders.sa_t5.SAT5Tokenizer
+        elif te_model in (TEModel.GEMMA_4_E4B, TEModel.GEMMA_4_E2B, TEModel.GEMMA_4_31B):
+            variant = {TEModel.GEMMA_4_E4B: comfy.text_encoders.gemma4.Gemma4_E4B,
+                       TEModel.GEMMA_4_E2B: comfy.text_encoders.gemma4.Gemma4_E2B,
+                       TEModel.GEMMA_4_31B: comfy.text_encoders.gemma4.Gemma4_31B}[te_model]
+            clip_target.clip = comfy.text_encoders.gemma4.gemma4_te(**llama_detect(clip_data), model_class=variant)
+            clip_target.tokenizer = variant.tokenizer
+            tokenizer_data["tokenizer_json"] = clip_data[0].get("tokenizer_json", None)
         elif te_model == TEModel.GEMMA_2_2B:
             clip_target.clip = comfy.text_encoders.lumina2.te(**llama_detect(clip_data))
             clip_target.tokenizer = comfy.text_encoders.lumina2.LuminaTokenizer

comfy/supported_models.py

@@ -27,6 +27,7 @@ import comfy.text_encoders.anima
 import comfy.text_encoders.ace15
 import comfy.text_encoders.longcat_image
 import comfy.text_encoders.ernie
+import comfy.text_encoders.cogvideo
 
 from . import supported_models_base
 from . import latent_formats
@@ -1781,6 +1782,183 @@ class ErnieImage(supported_models_base.BASE):
         return supported_models_base.ClipTarget(comfy.text_encoders.ernie.ErnieTokenizer, comfy.text_encoders.ernie.te(**hunyuan_detect))
 
 
-models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, LongCatImage, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImagePixelSpace, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, WAN21_FlowRVS, WAN21_SCAIL, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, ACEStep15, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima, RT_DETR_v4, ErnieImage]
+class SAM3(supported_models_base.BASE):
+    unet_config = {"image_model": "SAM3"}
+    supported_inference_dtypes = [torch.float16, torch.bfloat16, torch.float32]
+    text_encoder_key_prefix = ["detector.backbone.language_backbone."]
+    unet_extra_prefix = ""
 
-models += [SVD_img2vid]
+    def process_clip_state_dict(self, state_dict):
+        clip_keys = getattr(self, "_clip_stash", {})
+        clip_keys = utils.state_dict_prefix_replace(clip_keys, {"detector.backbone.language_backbone.": "", "backbone.language_backbone.": ""}, filter_keys=True)
+        clip_keys = utils.clip_text_transformers_convert(clip_keys, "encoder.", "sam3_clip.transformer.")
+        return {k: v for k, v in clip_keys.items() if not k.startswith("resizer.")}
+
+    def process_unet_state_dict(self, state_dict):
+        self._clip_stash = {k: state_dict.pop(k) for k in list(state_dict.keys()) if "language_backbone" in k and "resizer" not in k}
+        # SAM3.1: remap tracker.model.* -> tracker.*
+        for k in list(state_dict.keys()):
+            if k.startswith("tracker.model."):
+                state_dict["tracker." + k[len("tracker.model."):]] = state_dict.pop(k)
+        # SAM3.1: remove per-block freqs_cis buffers (computed dynamically)
+        for k in [k for k in list(state_dict.keys()) if ".attn.freqs_cis" in k]:
+            state_dict.pop(k)
+        # Split fused QKV projections
+        for k in [k for k in list(state_dict.keys()) if k.endswith((".in_proj_weight", ".in_proj_bias"))]:
+            t = state_dict.pop(k)
+            base, suffix = k.rsplit(".in_proj_", 1)
+            s = ".weight" if suffix == "weight" else ".bias"
+            d = t.shape[0] // 3
+            state_dict[base + ".q_proj" + s] = t[:d]
+            state_dict[base + ".k_proj" + s] = t[d:2*d]
+            state_dict[base + ".v_proj" + s] = t[2*d:]
+        # Remap tracker SAM decoder transformer key names to match sam.py TwoWayTransformer
+        for k in list(state_dict.keys()):
+            if "sam_mask_decoder.transformer." not in k:
+                continue
+            new_k = k.replace(".mlp.lin1.", ".mlp.0.").replace(".mlp.lin2.", ".mlp.2.").replace(".norm_final_attn.", ".norm_final.")
+            if new_k != k:
+                state_dict[new_k] = state_dict.pop(k)
+        return state_dict
+
+    def get_model(self, state_dict, prefix="", device=None):
+        return model_base.SAM3(self, device=device)
+
+    def clip_target(self, state_dict={}):
+        import comfy.text_encoders.sam3_clip
+        return supported_models_base.ClipTarget(comfy.text_encoders.sam3_clip.SAM3TokenizerWrapper, comfy.text_encoders.sam3_clip.SAM3ClipModelWrapper)
+
+
+class SAM31(SAM3):
+    unet_config = {"image_model": "SAM31"}
+
+
+class CogVideoX_T2V(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "cogvideox",
+    }
+
+    sampling_settings = {
+        "linear_start": 0.00085,
+        "linear_end": 0.012,
+        "beta_schedule": "linear",
+        "zsnr": True,
+    }
+
+    unet_extra_config = {}
+    latent_format = latent_formats.CogVideoX
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
+
+    vae_key_prefix = ["vae."]
+    text_encoder_key_prefix = ["text_encoders."]
+
+    def get_model(self, state_dict, prefix="", device=None):
+        # CogVideoX 1.5 (patch_size_t=2) has different training base dimensions for RoPE
+        if self.unet_config.get("patch_size_t") is not None:
+            self.unet_config.setdefault("sample_height", 96)
+            self.unet_config.setdefault("sample_width", 170)
+            self.unet_config.setdefault("sample_frames", 81)
+        out = model_base.CogVideoX(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        return supported_models_base.ClipTarget(comfy.text_encoders.cogvideo.CogVideoXT5Tokenizer, comfy.text_encoders.sd3_clip.T5XXLModel)
+
+class CogVideoX_I2V(CogVideoX_T2V):
+    unet_config = {
+        "image_model": "cogvideox",
+        "in_channels": 32,
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        if self.unet_config.get("patch_size_t") is not None:
+            self.unet_config.setdefault("sample_height", 96)
+            self.unet_config.setdefault("sample_width", 170)
+            self.unet_config.setdefault("sample_frames", 81)
+        out = model_base.CogVideoX(self, image_to_video=True, device=device)
+        return out
+
+
+models = [
+    LotusD,
+    Stable_Zero123,
+    SD15_instructpix2pix,
+    SD15,
+    SD20,
+    SD21UnclipL,
+    SD21UnclipH,
+    SDXL_instructpix2pix,
+    SDXLRefiner,
+    SDXL,
+    SSD1B,
+    KOALA_700M,
+    KOALA_1B,
+    Segmind_Vega,
+    SD_X4Upscaler,
+    Stable_Cascade_C,
+    Stable_Cascade_B,
+    SV3D_u,
+    SV3D_p,
+    SD3,
+    StableAudio,
+    AuraFlow,
+    PixArtAlpha,
+    PixArtSigma,
+    HunyuanDiT,
+    HunyuanDiT1,
+    FluxInpaint,
+    Flux,
+    LongCatImage,
+    FluxSchnell,
+    GenmoMochi,
+    LTXV,
+    LTXAV,
+    HunyuanVideo15_SR_Distilled,
+    HunyuanVideo15,
+    HunyuanImage21Refiner,
+    HunyuanImage21,
+    HunyuanVideoSkyreelsI2V,
+    HunyuanVideoI2V,
+    HunyuanVideo,
+    CosmosT2V,
+    CosmosI2V,
+    CosmosT2IPredict2,
+    CosmosI2VPredict2,
+    ZImagePixelSpace,
+    ZImage,
+    Lumina2,
+    WAN22_T2V,
+    WAN21_T2V,
+    WAN21_I2V,
+    WAN21_FunControl2V,
+    WAN21_Vace,
+    WAN21_Camera,
+    WAN22_Camera,
+    WAN22_S2V,
+    WAN21_HuMo,
+    WAN22_Animate,
+    WAN21_FlowRVS,
+    WAN21_SCAIL,
+    Hunyuan3Dv2mini,
+    Hunyuan3Dv2,
+    Hunyuan3Dv2_1,
+    HiDream,
+    Chroma,
+    ChromaRadiance,
+    ACEStep,
+    ACEStep15,
+    Omnigen2,
+    QwenImage,
+    Flux2,
+    Kandinsky5Image,
+    Kandinsky5,
+    Anima,
+    RT_DETR_v4,
+    ErnieImage,
+    SAM3,
+    SAM31,
+    CogVideoX_I2V,
+    CogVideoX_T2V,
+    SVD_img2vid,
+]

comfy/taesd/taehv.py

@@ -7,6 +7,7 @@ from tqdm.auto import tqdm
 from collections import namedtuple, deque
 
 import comfy.ops
+import comfy.model_management
 operations=comfy.ops.disable_weight_init
 
 DecoderResult = namedtuple("DecoderResult", ("frame", "memory"))
@@ -47,11 +48,14 @@ class TGrow(nn.Module):
         x = self.conv(x)
         return x.reshape(-1, C, H, W)
 
-def apply_model_with_memblocks(model, x, parallel, show_progress_bar):
+def apply_model_with_memblocks(model, x, parallel, show_progress_bar, output_device=None,
+                               patch_size=1, decode=False):
 
     B, T, C, H, W = x.shape
     if parallel:
         x = x.reshape(B*T, C, H, W)
+        if not decode and patch_size > 1:
+            x = F.pixel_unshuffle(x, patch_size)
         # parallel over input timesteps, iterate over blocks
         for b in tqdm(model, disable=not show_progress_bar):
             if isinstance(b, MemBlock):
@@ -62,20 +66,27 @@ def apply_model_with_memblocks(model, x, parallel, show_progress_bar):
                 x = b(x, mem)
             else:
                 x = b(x)
-        BT, C, H, W = x.shape
-        T = BT // B
-        x = x.view(B, T, C, H, W)
+        if decode and patch_size > 1:
+            x = F.pixel_shuffle(x, patch_size)
+        x = x.view(B, x.shape[0] // B, *x.shape[1:])
+        x = x.to(output_device)
     else:
         out = []
-        work_queue = deque([TWorkItem(xt, 0) for t, xt in enumerate(x.reshape(B, T * C, H, W).chunk(T, dim=1))])
+        # Chunk along the time dim directly (chunks are [B,1,C,H,W] views, squeeze to [B,C,H,W] views).
+        # Avoids forcing a contiguous copy when x is non-contiguous (e.g. after movedim in encode/decode).
+        work_queue = deque([TWorkItem(xt.squeeze(1), 0) for xt in x.chunk(T, dim=1)])
         progress_bar = tqdm(range(T), disable=not show_progress_bar)
         mem = [None] * len(model)
         while work_queue:
             xt, i = work_queue.popleft()
             if i == 0:
                 progress_bar.update(1)
+                if not decode and patch_size > 1:
+                    xt = F.pixel_unshuffle(xt, patch_size)
             if i == len(model):
-                out.append(xt)
+                if decode and patch_size > 1:
+                    xt = F.pixel_shuffle(xt, patch_size)
+                out.append(xt.to(output_device))
                 del xt
             else:
                 b = model[i]
@@ -165,24 +176,20 @@ class TAEHV(nn.Module):
 
     def encode(self, x, **kwargs):
         x = x.movedim(2, 1)  # [B, C, T, H, W] -> [B, T, C, H, W]
-        if self.patch_size > 1:
-            B, T, C, H, W = x.shape
-            x = x.reshape(B * T, C, H, W)
-            x = F.pixel_unshuffle(x, self.patch_size)
-            x = x.reshape(B, T, C * self.patch_size ** 2, H // self.patch_size, W // self.patch_size)
         if x.shape[1] % self.t_downscale != 0:
             # pad at end to multiple of t_downscale
             n_pad = self.t_downscale - x.shape[1] % self.t_downscale
             padding = x[:, -1:].repeat_interleave(n_pad, dim=1)
             x = torch.cat([x, padding], 1)
-        x = apply_model_with_memblocks(self.encoder, x, self.parallel, self.show_progress_bar).movedim(2, 1)
+        x = apply_model_with_memblocks(self.encoder, x, self.parallel, self.show_progress_bar,
+                                        patch_size=self.patch_size).movedim(2, 1)
         return self.process_out(x)
 
     def decode(self, x, **kwargs):
         x = x.unsqueeze(0) if x.ndim == 4 else x  # [T, C, H, W] -> [1, T, C, H, W]
         x = x.movedim(1, 2) if x.shape[1] != self.latent_channels else x  # [B, T, C, H, W] or [B, C, T, H, W]
         x = self.process_in(x).movedim(2, 1)  # [B, C, T, H, W] -> [B, T, C, H, W]
-        x = apply_model_with_memblocks(self.decoder, x, self.parallel, self.show_progress_bar)
-        if self.patch_size > 1:
-            x = F.pixel_shuffle(x, self.patch_size)
+        x = apply_model_with_memblocks(self.decoder, x, self.parallel, self.show_progress_bar,
+                                        output_device=comfy.model_management.intermediate_device(),
+                                        patch_size=self.patch_size, decode=True)
         return x[:, self.frames_to_trim:].movedim(2, 1)

comfy/taesd/taesd.py

@@ -17,32 +17,79 @@ class Clamp(nn.Module):
         return torch.tanh(x / 3) * 3
 
 class Block(nn.Module):
-    def __init__(self, n_in, n_out):
+    def __init__(self, n_in: int, n_out: int, use_midblock_gn: bool = False):
         super().__init__()
         self.conv = nn.Sequential(conv(n_in, n_out), nn.ReLU(), conv(n_out, n_out), nn.ReLU(), conv(n_out, n_out))
         self.skip = comfy.ops.disable_weight_init.Conv2d(n_in, n_out, 1, bias=False) if n_in != n_out else nn.Identity()
         self.fuse = nn.ReLU()
-    def forward(self, x):
+        if not use_midblock_gn:
+            self.pool = None
+            return
+        n_gn = n_in * 4
+        self.pool = nn.Sequential(
+            comfy.ops.disable_weight_init.Conv2d(n_in, n_gn, 1, bias=False),
+            comfy.ops.disable_weight_init.GroupNorm(4, n_gn),
+            nn.ReLU(inplace=True),
+            comfy.ops.disable_weight_init.Conv2d(n_gn, n_in, 1, bias=False),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if self.pool is not None:
+            x = x + self.pool(x)
         return self.fuse(self.conv(x) + self.skip(x))
 
-def Encoder(latent_channels=4):
-    return nn.Sequential(
-        conv(3, 64), Block(64, 64),
-        conv(64, 64, stride=2, bias=False), Block(64, 64), Block(64, 64), Block(64, 64),
-        conv(64, 64, stride=2, bias=False), Block(64, 64), Block(64, 64), Block(64, 64),
-        conv(64, 64, stride=2, bias=False), Block(64, 64), Block(64, 64), Block(64, 64),
-        conv(64, latent_channels),
-    )
+class Encoder(nn.Sequential):
+    def __init__(self, latent_channels: int = 4, use_gn: bool = False):
+        super().__init__(
+            conv(3, 64), Block(64, 64),
+            conv(64, 64, stride=2, bias=False), Block(64, 64), Block(64, 64), Block(64, 64),
+            conv(64, 64, stride=2, bias=False), Block(64, 64), Block(64, 64), Block(64, 64),
+            conv(64, 64, stride=2, bias=False), Block(64, 64, use_gn), Block(64, 64, use_gn), Block(64, 64, use_gn),
+            conv(64, latent_channels),
+        )
 
+class Decoder(nn.Sequential):
+    def __init__(self, latent_channels: int = 4, use_gn: bool = False):
+        super().__init__(
+            Clamp(), conv(latent_channels, 64), nn.ReLU(),
+            Block(64, 64, use_gn), Block(64, 64, use_gn), Block(64, 64, use_gn), nn.Upsample(scale_factor=2), conv(64, 64, bias=False),
+            Block(64, 64), Block(64, 64), Block(64, 64), nn.Upsample(scale_factor=2), conv(64, 64, bias=False),
+            Block(64, 64), Block(64, 64), Block(64, 64), nn.Upsample(scale_factor=2), conv(64, 64, bias=False),
+            Block(64, 64), conv(64, 3),
+        )
+
+class DecoderFlux2(Decoder):
+    def __init__(self, latent_channels: int = 128, use_gn: bool = True):
+        if latent_channels != 128 or not use_gn:
+            raise ValueError("Unexpected parameters for Flux2 TAE module")
+        super().__init__(latent_channels=32, use_gn=True)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        B, C, H, W = x.shape
+        x = (
+            x
+            .reshape(B, 32, 2, 2, H, W)
+            .permute(0, 1, 4, 2, 5, 3)
+            .reshape(B, 32, H * 2, W * 2)
+        )
+        return super().forward(x)
+
+class EncoderFlux2(Encoder):
+    def __init__(self, latent_channels: int = 128, use_gn: bool = True):
+        if latent_channels != 128 or not use_gn:
+            raise ValueError("Unexpected parameters for Flux2 TAE module")
+        super().__init__(latent_channels=32, use_gn=True)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        result = super().forward(x)
+        B, C, H, W = result.shape
+        return (
+            result
+            .reshape(B, C, H // 2, 2, W // 2, 2)
+            .permute(0, 1, 3, 5, 2, 4)
+            .reshape(B, 128, H // 2, W // 2)
+        )
 
-def Decoder(latent_channels=4):
-    return nn.Sequential(
-        Clamp(), conv(latent_channels, 64), nn.ReLU(),
-        Block(64, 64), Block(64, 64), Block(64, 64), nn.Upsample(scale_factor=2), conv(64, 64, bias=False),
-        Block(64, 64), Block(64, 64), Block(64, 64), nn.Upsample(scale_factor=2), conv(64, 64, bias=False),
-        Block(64, 64), Block(64, 64), Block(64, 64), nn.Upsample(scale_factor=2), conv(64, 64, bias=False),
-        Block(64, 64), conv(64, 3),
-    )
 
 class TAESD(nn.Module):
     latent_magnitude = 3
@@ -51,8 +98,15 @@ class TAESD(nn.Module):
     def __init__(self, encoder_path=None, decoder_path=None, latent_channels=4):
         """Initialize pretrained TAESD on the given device from the given checkpoints."""
         super().__init__()
-        self.taesd_encoder = Encoder(latent_channels=latent_channels)
-        self.taesd_decoder = Decoder(latent_channels=latent_channels)
+        if latent_channels == 128:
+            encoder_class = EncoderFlux2
+            decoder_class = DecoderFlux2
+        else:
+            encoder_class = Encoder
+            decoder_class = Decoder
+        self.taesd_encoder = encoder_class(latent_channels=latent_channels)
+        self.taesd_decoder = decoder_class(latent_channels=latent_channels)
+
         self.vae_scale = torch.nn.Parameter(torch.tensor(1.0))
         self.vae_shift = torch.nn.Parameter(torch.tensor(0.0))
         if encoder_path is not None:
@@ -61,19 +115,19 @@ class TAESD(nn.Module):
             self.taesd_decoder.load_state_dict(comfy.utils.load_torch_file(decoder_path, safe_load=True))
 
     @staticmethod
-    def scale_latents(x):
+    def scale_latents(x: torch.Tensor) -> torch.Tensor:
         """raw latents -> [0, 1]"""
         return x.div(2 * TAESD.latent_magnitude).add(TAESD.latent_shift).clamp(0, 1)
 
     @staticmethod
-    def unscale_latents(x):
+    def unscale_latents(x: torch.Tensor) -> torch.Tensor:
         """[0, 1] -> raw latents"""
         return x.sub(TAESD.latent_shift).mul(2 * TAESD.latent_magnitude)
 
-    def decode(self, x):
+    def decode(self, x: torch.Tensor) -> torch.Tensor:
         x_sample = self.taesd_decoder((x - self.vae_shift) * self.vae_scale)
         x_sample = x_sample.sub(0.5).mul(2)
         return x_sample
 
-    def encode(self, x):
+    def encode(self, x: torch.Tensor) -> torch.Tensor:
         return (self.taesd_encoder(x * 0.5 + 0.5) / self.vae_scale) + self.vae_shift

comfy/text_encoders/cogvideo.py

@@ -0,0 +1,6 @@
+import comfy.text_encoders.sd3_clip
+
+
+class CogVideoXT5Tokenizer(comfy.text_encoders.sd3_clip.T5XXLTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, min_length=226)

comfy/text_encoders/llama.py

@@ -521,7 +521,7 @@ class Attention(nn.Module):
             else:
                 present_key_value = (xk, xv, index + num_tokens)
 
-            if sliding_window is not None and xk.shape[2] > sliding_window:
+            if sliding_window is not None and xk.shape[2] > sliding_window and seq_length == 1:
                 xk = xk[:, :, -sliding_window:]
                 xv = xv[:, :, -sliding_window:]
                 attention_mask = attention_mask[..., -sliding_window:] if attention_mask is not None else None
@@ -533,12 +533,12 @@ class Attention(nn.Module):
         return self.o_proj(output), present_key_value
 
 class MLP(nn.Module):
-    def __init__(self, config: Llama2Config, device=None, dtype=None, ops: Any = None):
+    def __init__(self, config: Llama2Config, device=None, dtype=None, ops: Any = None, intermediate_size=None):
         super().__init__()
-        ops = ops or nn
-        self.gate_proj = ops.Linear(config.hidden_size, config.intermediate_size, bias=False, device=device, dtype=dtype)
-        self.up_proj = ops.Linear(config.hidden_size, config.intermediate_size, bias=False, device=device, dtype=dtype)
-        self.down_proj = ops.Linear(config.intermediate_size, config.hidden_size, bias=False, device=device, dtype=dtype)
+        intermediate_size = intermediate_size or config.intermediate_size
+        self.gate_proj = ops.Linear(config.hidden_size, intermediate_size, bias=False, device=device, dtype=dtype)
+        self.up_proj = ops.Linear(config.hidden_size, intermediate_size, bias=False, device=device, dtype=dtype)
+        self.down_proj = ops.Linear(intermediate_size, config.hidden_size, bias=False, device=device, dtype=dtype)
         if config.mlp_activation == "silu":
             self.activation = torch.nn.functional.silu
         elif config.mlp_activation == "gelu_pytorch_tanh":
@@ -647,24 +647,25 @@ class TransformerBlockGemma2(nn.Module):
 
         return x, present_key_value
 
+def _make_scaled_embedding(ops, vocab_size, hidden_size, scale, device, dtype):
+    class ScaledEmbedding(ops.Embedding):
+        def forward(self, input_ids, out_dtype=None):
+            return super().forward(input_ids, out_dtype=out_dtype) * scale
+    return ScaledEmbedding(vocab_size, hidden_size, device=device, dtype=dtype)
+
+
 class Llama2_(nn.Module):
     def __init__(self, config, device=None, dtype=None, ops=None):
         super().__init__()
         self.config = config
         self.vocab_size = config.vocab_size
 
-        self.embed_tokens = ops.Embedding(
-            config.vocab_size,
-            config.hidden_size,
-            device=device,
-            dtype=dtype
-        )
         if self.config.transformer_type == "gemma2" or self.config.transformer_type == "gemma3":
             transformer = TransformerBlockGemma2
-            self.normalize_in = True
+            self.embed_tokens = _make_scaled_embedding(ops, config.vocab_size, config.hidden_size, config.hidden_size ** 0.5, device, dtype)
         else:
             transformer = TransformerBlock
-            self.normalize_in = False
+            self.embed_tokens = ops.Embedding(config.vocab_size, config.hidden_size, device=device, dtype=dtype)
 
         self.layers = nn.ModuleList([
             transformer(config, index=i, device=device, dtype=dtype, ops=ops)
@@ -690,15 +691,12 @@ class Llama2_(nn.Module):
                                     self.config.rope_dims,
                                     device=device)
 
-    def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, position_ids=None, embeds_info=[], past_key_values=None):
+    def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, position_ids=None, embeds_info=[], past_key_values=None, input_ids=None):
         if embeds is not None:
             x = embeds
         else:
             x = self.embed_tokens(x, out_dtype=dtype)
 
-        if self.normalize_in:
-            x *= self.config.hidden_size ** 0.5
-
         seq_len = x.shape[1]
         past_len = 0
         if past_key_values is not None and len(past_key_values) > 0:
@@ -850,7 +848,7 @@ class BaseGenerate:
                                     torch.empty([batch, model_config.num_key_value_heads, max_cache_len, model_config.head_dim], device=device, dtype=execution_dtype), 0))
         return past_key_values
 
-    def generate(self, embeds=None, do_sample=True, max_length=256, temperature=1.0, top_k=50, top_p=0.9, min_p=0.0, repetition_penalty=1.0, seed=42, stop_tokens=None, initial_tokens=[], execution_dtype=None, min_tokens=0, presence_penalty=0.0):
+    def generate(self, embeds=None, do_sample=True, max_length=256, temperature=1.0, top_k=50, top_p=0.9, min_p=0.0, repetition_penalty=1.0, seed=42, stop_tokens=None, initial_tokens=[], execution_dtype=None, min_tokens=0, presence_penalty=0.0, initial_input_ids=None):
         device = embeds.device
 
         if stop_tokens is None:
@@ -875,14 +873,16 @@ class BaseGenerate:
         pbar = comfy.utils.ProgressBar(max_length)
 
         # Generation loop
+        current_input_ids = initial_input_ids
         for step in tqdm(range(max_length), desc="Generating tokens"):
-            x, _, past_key_values = self.model.forward(None, embeds=embeds, attention_mask=None, past_key_values=past_key_values)
+            x, _, past_key_values = self.model.forward(None, embeds=embeds, attention_mask=None, past_key_values=past_key_values, input_ids=current_input_ids)
             logits = self.logits(x)[:, -1]
             next_token = self.sample_token(logits, temperature, top_k, top_p, min_p, repetition_penalty, initial_tokens + generated_token_ids, generator, do_sample=do_sample, presence_penalty=presence_penalty)
             token_id = next_token[0].item()
             generated_token_ids.append(token_id)
 
             embeds = self.model.embed_tokens(next_token).to(execution_dtype)
+            current_input_ids = next_token if initial_input_ids is not None else None
             pbar.update(1)
 
             if token_id in stop_tokens:

comfy/text_encoders/lt.py

@@ -93,8 +93,7 @@ class Gemma3_12BModel(sd1_clip.SDClipModel):
 
     def generate(self, tokens, do_sample, max_length, temperature, top_k, top_p, min_p, repetition_penalty, seed, presence_penalty):
         tokens_only = [[t[0] for t in b] for b in tokens]
-        embeds, _, _, embeds_info = self.process_tokens(tokens_only, self.execution_device)
-        comfy.utils.normalize_image_embeddings(embeds, embeds_info, self.transformer.model.config.hidden_size ** 0.5)
+        embeds, _, _, _ = self.process_tokens(tokens_only, self.execution_device)
         return self.transformer.generate(embeds, do_sample, max_length, temperature, top_k, top_p, min_p, repetition_penalty, seed, stop_tokens=[106], presence_penalty=presence_penalty)  # 106 is <end_of_turn>
 
 class DualLinearProjection(torch.nn.Module):

comfy/text_encoders/lumina2.py

@@ -50,8 +50,7 @@ class Gemma3_4B_Vision_Model(sd1_clip.SDClipModel):
         super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma3_4B_Vision, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
 
     def process_tokens(self, tokens, device):
-        embeds, _, _, embeds_info = super().process_tokens(tokens, device)
-        comfy.utils.normalize_image_embeddings(embeds, embeds_info, self.transformer.model.config.hidden_size ** 0.5)
+        embeds, _, _, _ = super().process_tokens(tokens, device)
         return embeds
 
 class LuminaModel(sd1_clip.SD1ClipModel):

comfy/text_encoders/qwen35.py

@@ -408,8 +408,6 @@ class Qwen35Transformer(Llama2_):
         nn.Module.__init__(self)
         self.config = config
         self.vocab_size = config.vocab_size
-        self.normalize_in = False
-
         self.embed_tokens = ops.Embedding(config.vocab_size, config.hidden_size, device=device, dtype=dtype)
         self.layers = nn.ModuleList([
             Qwen35TransformerBlock(config, index=i, device=device, dtype=dtype, ops=ops)

comfy/text_encoders/sam3_clip.py

@@ -0,0 +1,97 @@
+import re
+from comfy import sd1_clip
+
+SAM3_CLIP_CONFIG = {
+    "architectures": ["CLIPTextModel"],
+    "hidden_act": "quick_gelu",
+    "hidden_size": 1024,
+    "intermediate_size": 4096,
+    "num_attention_heads": 16,
+    "num_hidden_layers": 24,
+    "max_position_embeddings": 32,
+    "projection_dim": 512,
+    "vocab_size": 49408,
+    "layer_norm_eps": 1e-5,
+    "eos_token_id": 49407,
+}
+
+
+class SAM3ClipModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, dtype=dtype, max_length=32, layer="last", textmodel_json_config=SAM3_CLIP_CONFIG, special_tokens={"start": 49406, "end": 49407, "pad": 0}, return_projected_pooled=False, return_attention_masks=True, enable_attention_masks=True, model_options=model_options)
+
+
+class SAM3Tokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(max_length=32, pad_with_end=False, pad_token=0, embedding_directory=embedding_directory, embedding_size=1024, embedding_key="sam3_clip", tokenizer_data=tokenizer_data)
+        self.disable_weights = True
+
+
+def _parse_prompts(text):
+    """Split comma-separated prompts with optional :N max detections per category"""
+    text = text.replace("(", "").replace(")", "")
+    parts = [p.strip() for p in text.split(",") if p.strip()]
+    result = []
+    for part in parts:
+        m = re.match(r'^(.+?)\s*:\s*([\d.]+)\s*$', part)
+        if m:
+            text_part = m.group(1).strip()
+            val = m.group(2)
+            max_det = max(1, round(float(val)))
+            result.append((text_part, max_det))
+        else:
+            result.append((part, 1))
+    return result
+
+
+class SAM3TokenizerWrapper(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="l", tokenizer=SAM3Tokenizer, name="sam3_clip")
+
+    def tokenize_with_weights(self, text: str, return_word_ids=False, **kwargs):
+        parsed = _parse_prompts(text)
+        if len(parsed) <= 1 and (not parsed or parsed[0][1] == 1):
+            return super().tokenize_with_weights(text, return_word_ids, **kwargs)
+        # Tokenize each prompt part separately, store per-part batches and metadata
+        inner = getattr(self, self.clip)
+        per_prompt = []
+        for prompt_text, max_det in parsed:
+            batches = inner.tokenize_with_weights(prompt_text, return_word_ids, **kwargs)
+            per_prompt.append((batches, max_det))
+        # Main output uses first prompt's tokens (for compatibility)
+        out = {self.clip_name: per_prompt[0][0], "sam3_per_prompt": per_prompt}
+        return out
+
+
+class SAM3ClipModelWrapper(sd1_clip.SD1ClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs):
+        super().__init__(device=device, dtype=dtype, model_options=model_options, clip_name="l", clip_model=SAM3ClipModel, name="sam3_clip")
+
+    def encode_token_weights(self, token_weight_pairs):
+        per_prompt = token_weight_pairs.pop("sam3_per_prompt", None)
+        if per_prompt is None:
+            return super().encode_token_weights(token_weight_pairs)
+
+        # Encode each prompt separately, pack into extra dict
+        inner = getattr(self, self.clip)
+        multi_cond = []
+        first_pooled = None
+        for batches, max_det in per_prompt:
+            out = inner.encode_token_weights(batches)
+            cond, pooled = out[0], out[1]
+            extra = out[2] if len(out) > 2 else {}
+            if first_pooled is None:
+                first_pooled = pooled
+            multi_cond.append({
+                "cond": cond,
+                "attention_mask": extra.get("attention_mask"),
+                "max_detections": max_det,
+            })
+
+        # Return first prompt as main (for non-SAM3 consumers), all prompts in metadata
+        main = multi_cond[0]
+        main_extra = {}
+        if main["attention_mask"] is not None:
+            main_extra["attention_mask"] = main["attention_mask"]
+        main_extra["sam3_multi_cond"] = multi_cond
+        return (main["cond"], first_pooled, main_extra)

comfy/utils.py

@@ -1446,10 +1446,3 @@ def deepcopy_list_dict(obj, memo=None):
     memo[obj_id] = res
     return res
 
-def normalize_image_embeddings(embeds, embeds_info, scale_factor):
-    """Normalize image embeddings to match text embedding scale"""
-    for info in embeds_info:
-        if info.get("type") == "image":
-            start_idx = info["index"]
-            end_idx = start_idx + info["size"]
-            embeds[:, start_idx:end_idx, :] /= scale_factor

comfy_api/input/__init__.py

@@ -9,6 +9,7 @@ from comfy_api.latest._input import (
     CurveInput,
     MonotoneCubicCurve,
     LinearCurve,
+    RangeInput,
 )
 
 __all__ = [
@@ -21,4 +22,5 @@ __all__ = [
     "CurveInput",
     "MonotoneCubicCurve",
     "LinearCurve",
+    "RangeInput",
 ]

comfy_api/latest/_input/__init__.py

@@ -1,5 +1,6 @@
 from .basic_types import ImageInput, AudioInput, MaskInput, LatentInput
 from .curve_types import CurvePoint, CurveInput, MonotoneCubicCurve, LinearCurve
+from .range_types import RangeInput
 from .video_types import VideoInput
 
 __all__ = [
@@ -12,4 +13,5 @@ __all__ = [
     "CurveInput",
     "MonotoneCubicCurve",
     "LinearCurve",
+    "RangeInput",
 ]

comfy_api/latest/_input/range_types.py

@@ -0,0 +1,70 @@
+from __future__ import annotations
+
+import logging
+import math
+import numpy as np
+
+logger = logging.getLogger(__name__)
+
+
+class RangeInput:
+    """Represents a levels/range adjustment: input range [min, max] with
+    optional midpoint (gamma control).
+
+    Generates a 1D LUT identical to GIMP's levels mapping:
+        1. Normalize input to [0, 1] using [min, max]
+        2. Apply gamma correction: pow(value, 1/gamma)
+        3. Clamp to [0, 1]
+
+    The midpoint field is a position in [0, 1] representing where the
+    midtone falls within [min, max]. It maps to gamma via:
+        gamma = -log2(midpoint)
+    So midpoint=0.5 → gamma=1.0 (linear).
+    """
+
+    def __init__(self, min_val: float, max_val: float, midpoint: float | None = None):
+        self.min_val = min_val
+        self.max_val = max_val
+        self.midpoint = midpoint
+
+    @staticmethod
+    def from_raw(data) -> RangeInput:
+        if isinstance(data, RangeInput):
+            return data
+        if isinstance(data, dict):
+            return RangeInput(
+                min_val=float(data.get("min", 0.0)),
+                max_val=float(data.get("max", 1.0)),
+                midpoint=float(data["midpoint"]) if data.get("midpoint") is not None else None,
+            )
+        raise TypeError(f"Cannot convert {type(data)} to RangeInput")
+
+    def to_lut(self, size: int = 256) -> np.ndarray:
+        """Generate a float64 lookup table mapping [0, 1] input through this
+        levels adjustment.
+
+        The LUT maps normalized input values (0..1) to output values (0..1),
+        matching the GIMP levels formula.
+        """
+        xs = np.linspace(0.0, 1.0, size, dtype=np.float64)
+
+        in_range = self.max_val - self.min_val
+        if abs(in_range) < 1e-10:
+            return np.where(xs >= self.min_val, 1.0, 0.0).astype(np.float64)
+
+        # Normalize: map [min, max] → [0, 1]
+        result = (xs - self.min_val) / in_range
+        result = np.clip(result, 0.0, 1.0)
+
+        # Gamma correction from midpoint
+        if self.midpoint is not None and self.midpoint > 0 and self.midpoint != 0.5:
+            gamma = max(-math.log2(self.midpoint), 0.001)
+            inv_gamma = 1.0 / gamma
+            mask = result > 0
+            result[mask] = np.power(result[mask], inv_gamma)
+
+        return result
+
+    def __repr__(self) -> str:
+        mid = f", midpoint={self.midpoint}" if self.midpoint is not None else ""
+        return f"RangeInput(min={self.min_val}, max={self.max_val}{mid})"

comfy_api/latest/_input_impl/video_types.py

@@ -12,6 +12,7 @@ import numpy as np
 import math
 import torch
 from .._util import VideoContainer, VideoCodec, VideoComponents
+import logging
 
 
 def container_to_output_format(container_format: str | None) -> str | None:
@@ -238,64 +239,125 @@ class VideoFromFile(VideoInput):
             start_time = max(self._get_raw_duration() + self.__start_time, 0)
         else:
             start_time = self.__start_time
+
         # Get video frames
         frames = []
+        audio_frames = []
+        alphas = None
         start_pts = int(start_time / video_stream.time_base)
         end_pts = int((start_time + self.__duration) / video_stream.time_base)
-        container.seek(start_pts, stream=video_stream)
-        for frame in container.decode(video_stream):
-            if frame.pts < start_pts:
-                continue
-            if self.__duration and frame.pts >= end_pts:
-                break
-            img = frame.to_ndarray(format='rgb24')  # shape: (H, W, 3)
-            img = torch.from_numpy(img) / 255.0  # shape: (H, W, 3)
-            frames.append(img)
 
-        images = torch.stack(frames) if len(frames) > 0 else torch.zeros(0, 3, 0, 0)
+        if start_pts != 0:
+            container.seek(start_pts, stream=video_stream)
+
+        image_format = 'gbrpf32le'
+        process_image_format = lambda a: a
+        audio = None
+
+        streams = [video_stream]
+        has_first_audio_frame = False
+        checked_alpha = False
+
+        # Default to False so we decode until EOF if duration is 0
+        video_done = False
+        audio_done = True
+
+        if len(container.streams.audio):
+            audio_stream = container.streams.audio[-1]
+            streams += [audio_stream]
+            resampler = av.audio.resampler.AudioResampler(format='fltp')
+            audio_done = False
+
+        for packet in container.demux(*streams):
+            if video_done and audio_done:
+                break
+
+            if packet.stream.type == "video":
+                if video_done:
+                    continue
+                try:
+                    for frame in packet.decode():
+                        if frame.pts < start_pts:
+                            continue
+                        if self.__duration and frame.pts >= end_pts:
+                            video_done = True
+                            break
+
+                        if not checked_alpha:
+                            alpha_channel = False
+                            for comp in frame.format.components:
+                                if comp.is_alpha or frame.format.name == "pal8":
+                                    alphas = []
+                                    alpha_channel = True
+                                    break
+                            if frame.format.name in ("yuvj420p", "yuvj422p", "yuvj444p", "rgb24", "rgba", "pal8"):
+                                process_image_format = lambda a: a.float() / 255.0
+                                if alpha_channel:
+                                    image_format = 'rgba'
+                                else:
+                                    image_format = 'rgb24'
+                            else:
+                                process_image_format = lambda a: a
+                                if alpha_channel:
+                                    image_format = 'gbrapf32le'
+                                else:
+                                    image_format = 'gbrpf32le'
+
+                            checked_alpha = True
+
+                        img = frame.to_ndarray(format=image_format)  # shape: (H, W, 4)
+                        if frame.rotation != 0:
+                            k = int(round(frame.rotation // 90))
+                            img = np.rot90(img, k=k, axes=(0, 1)).copy()
+                        if alphas is None:
+                            frames.append(torch.from_numpy(img))
+                        else:
+                            frames.append(torch.from_numpy(img[..., :-1]))
+                            alphas.append(torch.from_numpy(img[..., -1:]))
+                except av.error.InvalidDataError:
+                    logging.info("pyav decode error")
+
+            elif packet.stream.type == "audio":
+                if audio_done:
+                    continue
+
+                aframes = itertools.chain.from_iterable(
+                    map(resampler.resample, packet.decode())
+                )
+                for frame in aframes:
+                    if self.__duration and frame.time > start_time + self.__duration:
+                        audio_done = True
+                        break
+
+                    if not has_first_audio_frame:
+                        offset_seconds = start_time - frame.pts * audio_stream.time_base
+                        to_skip = max(0, int(offset_seconds * audio_stream.sample_rate))
+                        if to_skip < frame.samples:
+                            has_first_audio_frame = True
+                            audio_frames.append(frame.to_ndarray()[..., to_skip:])
+                    else:
+                        audio_frames.append(frame.to_ndarray())
+
+        images = process_image_format(torch.stack(frames)) if len(frames) > 0 else torch.zeros(0, 0, 0, 3)
+        if alphas is not None:
+            alphas = process_image_format(torch.stack(alphas)) if len(alphas) > 0 else torch.zeros(0, 0, 0, 1)
 
         # Get frame rate
         frame_rate = Fraction(video_stream.average_rate) if video_stream.average_rate else Fraction(1)
 
-        # Get audio if available
-        audio = None
-        container.seek(start_pts, stream=video_stream)
-        # Use last stream for consistency
-        if len(container.streams.audio):
-            audio_stream = container.streams.audio[-1]
-            audio_frames = []
-            resample = av.audio.resampler.AudioResampler(format='fltp').resample
-            frames = itertools.chain.from_iterable(
-                map(resample, container.decode(audio_stream))
-            )
+        if len(audio_frames) > 0:
+            audio_data = np.concatenate(audio_frames, axis=1)  # shape: (channels, total_samples)
+            if self.__duration:
+                audio_data = audio_data[..., :int(self.__duration * audio_stream.sample_rate)]
 
-            has_first_frame = False
-            for frame in frames:
-                offset_seconds = start_time - frame.pts * audio_stream.time_base
-                to_skip = max(0, int(offset_seconds * audio_stream.sample_rate))
-                if to_skip < frame.samples:
-                    has_first_frame = True
-                    break
-            if has_first_frame:
-                audio_frames.append(frame.to_ndarray()[..., to_skip:])
-
-            for frame in frames:
-                if self.__duration and frame.time > start_time + self.__duration:
-                    break
-                audio_frames.append(frame.to_ndarray())  # shape: (channels, samples)
-            if len(audio_frames) > 0:
-                audio_data = np.concatenate(audio_frames, axis=1)  # shape: (channels, total_samples)
-                if self.__duration:
-                    audio_data = audio_data[..., :int(self.__duration * audio_stream.sample_rate)]
-
-                audio_tensor = torch.from_numpy(audio_data).unsqueeze(0)  # shape: (1, channels, total_samples)
-                audio = AudioInput({
-                    "waveform": audio_tensor,
-                    "sample_rate": int(audio_stream.sample_rate) if audio_stream.sample_rate else 1,
-                })
+            audio_tensor = torch.from_numpy(audio_data).unsqueeze(0)  # shape: (1, channels, total_samples)
+            audio = AudioInput({
+                "waveform": audio_tensor,
+                "sample_rate": int(audio_stream.sample_rate) if audio_stream.sample_rate else 1,
+            })
 
         metadata = container.metadata
-        return VideoComponents(images=images, audio=audio, frame_rate=frame_rate, metadata=metadata)
+        return VideoComponents(images=images, alpha=alphas, audio=audio, frame_rate=frame_rate, metadata=metadata)
 
     def get_components(self) -> VideoComponents:
         if isinstance(self.__file, io.BytesIO):

comfy_api/latest/_io.py

@@ -1266,6 +1266,43 @@ class Histogram(ComfyTypeIO):
     Type = list[int]
 
 
+@comfytype(io_type="RANGE")
+class Range(ComfyTypeIO):
+    from comfy_api.input import RangeInput
+    if TYPE_CHECKING:
+        Type = RangeInput
+
+    class Input(WidgetInput):
+        def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None,
+                     socketless: bool=True, default: dict=None,
+                     display: str=None,
+                     gradient_stops: list=None,
+                     show_midpoint: bool=None,
+                     midpoint_scale: str=None,
+                     value_min: float=None,
+                     value_max: float=None,
+                     advanced: bool=None):
+            super().__init__(id, display_name, optional, tooltip, None, default, socketless, None, None, None, None, advanced)
+            if default is None:
+                self.default = {"min": 0.0, "max": 1.0}
+            self.display = display
+            self.gradient_stops = gradient_stops
+            self.show_midpoint = show_midpoint
+            self.midpoint_scale = midpoint_scale
+            self.value_min = value_min
+            self.value_max = value_max
+
+        def as_dict(self):
+            return super().as_dict() | prune_dict({
+                "display": self.display,
+                "gradient_stops": self.gradient_stops,
+                "show_midpoint": self.show_midpoint,
+                "midpoint_scale": self.midpoint_scale,
+                "value_min": self.value_min,
+                "value_max": self.value_max,
+            })
+
+
 DYNAMIC_INPUT_LOOKUP: dict[str, Callable[[dict[str, Any], dict[str, Any], tuple[str, dict[str, Any]], str, list[str] | None], None]] = {}
 def register_dynamic_input_func(io_type: str, func: Callable[[dict[str, Any], dict[str, Any], tuple[str, dict[str, Any]], str, list[str] | None], None]):
     DYNAMIC_INPUT_LOOKUP[io_type] = func
@@ -2276,5 +2313,6 @@ __all__ = [
     "BoundingBox",
     "Curve",
     "Histogram",
+    "Range",
     "NodeReplace",
 ]

comfy_api/latest/_util/video_types.py

@@ -3,7 +3,7 @@ from dataclasses import dataclass
 from enum import Enum
 from fractions import Fraction
 from typing import Optional
-from .._input import ImageInput, AudioInput
+from .._input import ImageInput, AudioInput, MaskInput
 
 class VideoCodec(str, Enum):
     AUTO = "auto"
@@ -48,5 +48,4 @@ class VideoComponents:
     frame_rate: Fraction
     audio: Optional[AudioInput] = None
     metadata: Optional[dict] = None
-
-
+    alpha: Optional[MaskInput] = None

comfy_api_nodes/apis/bytedance.py

@@ -122,6 +122,46 @@ class TaskStatusResponse(BaseModel):
     usage: TaskStatusUsage | None = Field(None)
 
 
+class GetAssetResponse(BaseModel):
+    id: str = Field(...)
+    name: str | None = Field(None)
+    url: str | None = Field(None)
+    asset_type: str = Field(...)
+    group_id: str = Field(...)
+    status: str = Field(...)
+    error: TaskStatusError | None = Field(None)
+
+
+class SeedanceCreateVisualValidateSessionResponse(BaseModel):
+    session_id: str = Field(...)
+    h5_link: str = Field(...)
+
+
+class SeedanceGetVisualValidateSessionResponse(BaseModel):
+    session_id: str = Field(...)
+    status: str = Field(...)
+    group_id: str | None = Field(None)
+    error_code: str | None = Field(None)
+    error_message: str | None = Field(None)
+
+
+class SeedanceCreateAssetRequest(BaseModel):
+    group_id: str = Field(...)
+    url: str = Field(...)
+    asset_type: str = Field(...)
+    name: str | None = Field(None, max_length=64)
+    project_name: str | None = Field(None)
+
+
+class SeedanceCreateAssetResponse(BaseModel):
+    asset_id: str = Field(...)
+
+
+class SeedanceVirtualLibraryCreateAssetRequest(BaseModel):
+    url: str = Field(..., description="Publicly accessible URL of the image asset to upload.")
+    hash: str = Field(..., description="Dedup key. Re-submitting the same hash returns the existing asset id.")
+
+
 # Dollars per 1K tokens, keyed by (model_id, has_video_input).
 SEEDANCE2_PRICE_PER_1K_TOKENS = {
     ("dreamina-seedance-2-0-260128", False): 0.007,

comfy_api_nodes/apis/moonvalley.py

@@ -1,152 +0,0 @@
-from enum import Enum
-from typing import Optional, Dict, Any
-
-from pydantic import BaseModel, Field, StrictBytes
-
-
-class MoonvalleyPromptResponse(BaseModel):
-    error: Optional[Dict[str, Any]] = None
-    frame_conditioning: Optional[Dict[str, Any]] = None
-    id: Optional[str] = None
-    inference_params: Optional[Dict[str, Any]] = None
-    meta: Optional[Dict[str, Any]] = None
-    model_params: Optional[Dict[str, Any]] = None
-    output_url: Optional[str] = None
-    prompt_text: Optional[str] = None
-    status: Optional[str] = None
-
-
-class MoonvalleyTextToVideoInferenceParams(BaseModel):
-    add_quality_guidance: Optional[bool] = Field(
-        True, description='Whether to add quality guidance'
-    )
-    caching_coefficient: Optional[float] = Field(
-        0.3, description='Caching coefficient for optimization'
-    )
-    caching_cooldown: Optional[int] = Field(
-        3, description='Number of caching cooldown steps'
-    )
-    caching_warmup: Optional[int] = Field(
-        3, description='Number of caching warmup steps'
-    )
-    clip_value: Optional[float] = Field(
-        3, description='CLIP value for generation control'
-    )
-    conditioning_frame_index: Optional[int] = Field(
-        0, description='Index of the conditioning frame'
-    )
-    cooldown_steps: Optional[int] = Field(
-        75, description='Number of cooldown steps (calculated based on num_frames)'
-    )
-    fps: Optional[int] = Field(
-        24, description='Frames per second of the generated video'
-    )
-    guidance_scale: Optional[float] = Field(
-        10, description='Guidance scale for generation control'
-    )
-    height: Optional[int] = Field(
-        1080, description='Height of the generated video in pixels'
-    )
-    negative_prompt: Optional[str] = Field(None, description='Negative prompt text')
-    num_frames: Optional[int] = Field(64, description='Number of frames to generate')
-    seed: Optional[int] = Field(
-        None, description='Random seed for generation (default: random)'
-    )
-    shift_value: Optional[float] = Field(
-        3, description='Shift value for generation control'
-    )
-    steps: Optional[int] = Field(80, description='Number of denoising steps')
-    use_guidance_schedule: Optional[bool] = Field(
-        True, description='Whether to use guidance scheduling'
-    )
-    use_negative_prompts: Optional[bool] = Field(
-        False, description='Whether to use negative prompts'
-    )
-    use_timestep_transform: Optional[bool] = Field(
-        True, description='Whether to use timestep transformation'
-    )
-    warmup_steps: Optional[int] = Field(
-        0, description='Number of warmup steps (calculated based on num_frames)'
-    )
-    width: Optional[int] = Field(
-        1920, description='Width of the generated video in pixels'
-    )
-
-
-class MoonvalleyTextToVideoRequest(BaseModel):
-    image_url: Optional[str] = None
-    inference_params: Optional[MoonvalleyTextToVideoInferenceParams] = None
-    prompt_text: Optional[str] = None
-    webhook_url: Optional[str] = None
-
-
-class MoonvalleyUploadFileRequest(BaseModel):
-    file: Optional[StrictBytes] = None
-
-
-class MoonvalleyUploadFileResponse(BaseModel):
-    access_url: Optional[str] = None
-
-
-class MoonvalleyVideoToVideoInferenceParams(BaseModel):
-    add_quality_guidance: Optional[bool] = Field(
-        True, description='Whether to add quality guidance'
-    )
-    caching_coefficient: Optional[float] = Field(
-        0.3, description='Caching coefficient for optimization'
-    )
-    caching_cooldown: Optional[int] = Field(
-        3, description='Number of caching cooldown steps'
-    )
-    caching_warmup: Optional[int] = Field(
-        3, description='Number of caching warmup steps'
-    )
-    clip_value: Optional[float] = Field(
-        3, description='CLIP value for generation control'
-    )
-    conditioning_frame_index: Optional[int] = Field(
-        0, description='Index of the conditioning frame'
-    )
-    cooldown_steps: Optional[int] = Field(
-        36, description='Number of cooldown steps (calculated based on num_frames)'
-    )
-    guidance_scale: Optional[float] = Field(
-        15, description='Guidance scale for generation control'
-    )
-    negative_prompt: Optional[str] = Field(None, description='Negative prompt text')
-    seed: Optional[int] = Field(
-        None, description='Random seed for generation (default: random)'
-    )
-    shift_value: Optional[float] = Field(
-        3, description='Shift value for generation control'
-    )
-    steps: Optional[int] = Field(80, description='Number of denoising steps')
-    use_guidance_schedule: Optional[bool] = Field(
-        True, description='Whether to use guidance scheduling'
-    )
-    use_negative_prompts: Optional[bool] = Field(
-        False, description='Whether to use negative prompts'
-    )
-    use_timestep_transform: Optional[bool] = Field(
-        True, description='Whether to use timestep transformation'
-    )
-    warmup_steps: Optional[int] = Field(
-        24, description='Number of warmup steps (calculated based on num_frames)'
-    )
-
-
-class ControlType(str, Enum):
-    motion_control = 'motion_control'
-    pose_control = 'pose_control'
-
-
-class MoonvalleyVideoToVideoRequest(BaseModel):
-    control_type: ControlType = Field(
-        ..., description='Supported types for video control'
-    )
-    inference_params: Optional[MoonvalleyVideoToVideoInferenceParams] = None
-    prompt_text: str = Field(..., description='Describes the video to generate')
-    video_url: str = Field(..., description='Url to control video')
-    webhook_url: Optional[str] = Field(
-        None, description='Optional webhook URL for notifications'
-    )

comfy_api_nodes/apis/topaz.py

@@ -1,4 +1,4 @@
-from typing import Optional, Union
+from typing import Optional
 
 from pydantic import BaseModel, Field
 
@@ -72,8 +72,11 @@ class VideoEnhancementFilter(BaseModel):
     grain: Optional[float] = Field(None, description="Grain after AI model processing")
     grainSize: Optional[float] = Field(None, description="Size of generated grain")
     recoverOriginalDetailValue: Optional[float] = Field(None, description="Source details into the output video")
-    creativity: Optional[str] = Field(None, description="Creativity level(high, low) for slc-1 only")
+    creativity: float | str | None = Field(None, description="slc-1/slp-2.5: enum (low/middle/high). ast-2: decimal 0.0-1.0.")
     isOptimizedMode: Optional[bool] = Field(None, description="Set to true for Starlight Creative (slc-1) only")
+    prompt: str | None = Field(None, description="Descriptive scene prompt (ast-2 only)")
+    sharp: float | None = Field(None, description="ast-2 pre-enhance sharpness")
+    realism: float | None = Field(None, description="ast-2 realism control")
 
 
 class OutputInformationVideo(BaseModel):
@@ -90,7 +93,7 @@ class Overrides(BaseModel):
 
 class CreateVideoRequest(BaseModel):
     source: CreateVideoRequestSource = Field(...)
-    filters: list[Union[VideoFrameInterpolationFilter, VideoEnhancementFilter]] = Field(...)
+    filters: list[VideoFrameInterpolationFilter | VideoEnhancementFilter] = Field(...)
     output: OutputInformationVideo = Field(...)
     overrides: Overrides = Field(Overrides(isPaidDiffusion=True))
 

comfy_api_nodes/apis/wan.py

@@ -118,7 +118,7 @@ class Wan27ReferenceVideoInputField(BaseModel):
 class Wan27ReferenceVideoParametersField(BaseModel):
     resolution: str = Field(...)
     ratio: str | None = Field(None)
-    duration: int = Field(5, ge=2, le=10)
+    duration: int = Field(5, ge=2, le=15)
     watermark: bool = Field(False)
     seed: int = Field(..., ge=0, le=2147483647)
 
@@ -157,7 +157,7 @@ class Wan27VideoEditInputField(BaseModel):
 class Wan27VideoEditParametersField(BaseModel):
     resolution: str = Field(...)
     ratio: str | None = Field(None)
-    duration: int = Field(0)
+    duration: int | None = Field(0)
     audio_setting: str = Field("auto")
     watermark: bool = Field(False)
     seed: int = Field(..., ge=0, le=2147483647)

comfy_api_nodes/nodes_bytedance.py

@@ -1,5 +1,7 @@
+import hashlib
 import logging
 import math
+import re
 
 import torch
 from typing_extensions import override
@@ -11,9 +13,15 @@ from comfy_api_nodes.apis.bytedance import (
     SEEDANCE2_PRICE_PER_1K_TOKENS,
     SEEDANCE2_REF_VIDEO_PIXEL_LIMITS,
     VIDEO_TASKS_EXECUTION_TIME,
+    GetAssetResponse,
     Image2VideoTaskCreationRequest,
     ImageTaskCreationResponse,
     Seedance2TaskCreationRequest,
+    SeedanceCreateAssetRequest,
+    SeedanceCreateAssetResponse,
+    SeedanceCreateVisualValidateSessionResponse,
+    SeedanceGetVisualValidateSessionResponse,
+    SeedanceVirtualLibraryCreateAssetRequest,
     Seedream4Options,
     Seedream4TaskCreationRequest,
     TaskAudioContent,
@@ -44,10 +52,16 @@ from comfy_api_nodes.util import (
     validate_image_aspect_ratio,
     validate_image_dimensions,
     validate_string,
+    validate_video_dimensions,
+    validate_video_duration,
 )
+from server import PromptServer
 
 BYTEPLUS_IMAGE_ENDPOINT = "/proxy/byteplus/api/v3/images/generations"
 
+_VERIFICATION_POLL_TIMEOUT_SEC = 120
+_VERIFICATION_POLL_INTERVAL_SEC = 3
+
 SEEDREAM_MODELS = {
     "seedream 5.0 lite": "seedream-5-0-260128",
     "seedream-4-5-251128": "seedream-4-5-251128",
@@ -96,6 +110,193 @@ def _validate_ref_video_pixels(video: Input.Video, model_id: str, resolution: st
         )
 
 
+async def _resolve_reference_assets(
+    cls: type[IO.ComfyNode],
+    asset_ids: list[str],
+) -> tuple[dict[str, str], dict[str, str], dict[str, str]]:
+    """Look up each asset, validate Active status, group by asset_type.
+
+    Returns (image_assets, video_assets, audio_assets), each mapping asset_id -> "asset://<asset_id>".
+    """
+    image_assets: dict[str, str] = {}
+    video_assets: dict[str, str] = {}
+    audio_assets: dict[str, str] = {}
+    for i, raw_id in enumerate(asset_ids, 1):
+        asset_id = (raw_id or "").strip()
+        if not asset_id:
+            continue
+        result = await sync_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/seedance/assets/{asset_id}"),
+            response_model=GetAssetResponse,
+        )
+        if result.status != "Active":
+            extra = f" {result.error.code}: {result.error.message}" if result.error else ""
+            raise ValueError(f"Reference asset {i} (Id={asset_id}) is not Active (Status={result.status}).{extra}")
+        asset_uri = f"asset://{asset_id}"
+        if result.asset_type == "Image":
+            image_assets[asset_id] = asset_uri
+        elif result.asset_type == "Video":
+            video_assets[asset_id] = asset_uri
+        elif result.asset_type == "Audio":
+            audio_assets[asset_id] = asset_uri
+    return image_assets, video_assets, audio_assets
+
+
+_ASSET_REF_RE = re.compile(r"\basset ?(\d{1,2})\b", re.IGNORECASE)
+
+
+def _build_asset_labels(
+    reference_assets: dict[str, str],
+    image_asset_uris: dict[str, str],
+    video_asset_uris: dict[str, str],
+    audio_asset_uris: dict[str, str],
+    n_reference_images: int,
+    n_reference_videos: int,
+    n_reference_audios: int,
+) -> dict[int, str]:
+    """Map asset slot number (from 'asset_N' keys) to its positional label.
+
+    Asset entries are appended to `content` after the reference_images/videos/audios,
+    so their 1-indexed labels continue from the count of existing same-type refs:
+    one reference_images entry + one Image-type asset -> asset labelled "Image 2".
+    """
+    image_n = n_reference_images
+    video_n = n_reference_videos
+    audio_n = n_reference_audios
+    labels: dict[int, str] = {}
+    for slot_key, raw_id in reference_assets.items():
+        asset_id = (raw_id or "").strip()
+        if not asset_id:
+            continue
+        try:
+            slot_num = int(slot_key.rsplit("_", 1)[-1])
+        except ValueError:
+            continue
+        if asset_id in image_asset_uris:
+            image_n += 1
+            labels[slot_num] = f"Image {image_n}"
+        elif asset_id in video_asset_uris:
+            video_n += 1
+            labels[slot_num] = f"Video {video_n}"
+        elif asset_id in audio_asset_uris:
+            audio_n += 1
+            labels[slot_num] = f"Audio {audio_n}"
+    return labels
+
+
+def _rewrite_asset_refs(prompt: str, labels: dict[int, str]) -> str:
+    """Case-insensitively replace 'assetNN' (1-2 digit) tokens with their labels."""
+    if not labels:
+        return prompt
+
+    def _sub(m: "re.Match[str]") -> str:
+        return labels.get(int(m.group(1)), m.group(0))
+
+    return _ASSET_REF_RE.sub(_sub, prompt)
+
+
+async def _obtain_group_id_via_h5_auth(cls: type[IO.ComfyNode]) -> str:
+    session = await sync_op(
+        cls,
+        ApiEndpoint(path="/proxy/seedance/visual-validate/sessions", method="POST"),
+        response_model=SeedanceCreateVisualValidateSessionResponse,
+    )
+    logger.warning("Seedance authentication required. Open link: %s", session.h5_link)
+
+    h5_text = f"Open this link in your browser and complete face verification:\n\n{session.h5_link}"
+
+    result = await poll_op(
+        cls,
+        ApiEndpoint(path=f"/proxy/seedance/visual-validate/sessions/{session.session_id}"),
+        response_model=SeedanceGetVisualValidateSessionResponse,
+        status_extractor=lambda r: r.status,
+        completed_statuses=["completed"],
+        failed_statuses=["failed"],
+        poll_interval=_VERIFICATION_POLL_INTERVAL_SEC,
+        max_poll_attempts=(_VERIFICATION_POLL_TIMEOUT_SEC // _VERIFICATION_POLL_INTERVAL_SEC) - 1,
+        estimated_duration=_VERIFICATION_POLL_TIMEOUT_SEC - 1,
+        extra_text=h5_text,
+    )
+
+    if not result.group_id:
+        raise RuntimeError(f"Seedance session {session.session_id} completed without a group_id")
+
+    logger.warning("Seedance authentication complete. New GroupId: %s", result.group_id)
+    PromptServer.instance.send_progress_text(
+        f"Authentication complete. New GroupId: {result.group_id}", cls.hidden.unique_id
+    )
+    return result.group_id
+
+
+async def _resolve_group_id(cls: type[IO.ComfyNode], group_id: str) -> str:
+    if group_id and group_id.strip():
+        return group_id.strip()
+    return await _obtain_group_id_via_h5_auth(cls)
+
+
+async def _create_seedance_asset(
+    cls: type[IO.ComfyNode],
+    *,
+    group_id: str,
+    url: str,
+    name: str,
+    asset_type: str,
+) -> str:
+    req = SeedanceCreateAssetRequest(
+        group_id=group_id,
+        url=url,
+        asset_type=asset_type,
+        name=name or None,
+    )
+    result = await sync_op(
+        cls,
+        ApiEndpoint(path="/proxy/seedance/assets", method="POST"),
+        response_model=SeedanceCreateAssetResponse,
+        data=req,
+    )
+    return result.asset_id
+
+
+async def _wait_for_asset_active(cls: type[IO.ComfyNode], asset_id: str, group_id: str) -> GetAssetResponse:
+    """Poll the newly created asset until its status becomes Active."""
+    return await poll_op(
+        cls,
+        ApiEndpoint(path=f"/proxy/seedance/assets/{asset_id}"),
+        response_model=GetAssetResponse,
+        status_extractor=lambda r: r.status,
+        completed_statuses=["Active"],
+        failed_statuses=["Failed"],
+        poll_interval=5,
+        max_poll_attempts=1200,
+        extra_text=f"Waiting for asset pre-processing...\n\nasset_id: {asset_id}\n\ngroup_id: {group_id}",
+    )
+
+
+async def _seedance_virtual_library_upload_image_asset(
+    cls: type[IO.ComfyNode],
+    image: torch.Tensor,
+    *,
+    wait_label: str = "Uploading image",
+) -> str:
+    """Upload an image into the caller's per-customer Seedance virtual library."""
+    public_url = await upload_image_to_comfyapi(cls, image, wait_label=wait_label)
+    normalized = image.detach().cpu().contiguous().to(torch.float32)
+    digest = hashlib.sha256()
+    digest.update(str(tuple(normalized.shape)).encode("utf-8"))
+    digest.update(b"\0")
+    digest.update(normalized.numpy().tobytes())
+    image_hash = digest.hexdigest()
+    create_resp = await sync_op(
+        cls,
+        ApiEndpoint(path="/proxy/seedance/virtual-library/assets", method="POST"),
+        response_model=SeedanceCreateAssetResponse,
+        data=SeedanceVirtualLibraryCreateAssetRequest(url=public_url, hash=image_hash),
+    )
+    await _wait_for_asset_active(cls, create_resp.asset_id, group_id="virtual-library")
+    return f"asset://{create_resp.asset_id}"
+
+
 def _seedance2_price_extractor(model_id: str, has_video_input: bool):
     """Returns a price_extractor closure for Seedance 2.0 poll_op."""
     rate = SEEDANCE2_PRICE_PER_1K_TOKENS.get((model_id, has_video_input))
@@ -1202,7 +1403,6 @@ class ByteDance2TextToVideoNode(IO.ComfyNode):
             status_extractor=lambda r: r.status,
             price_extractor=_seedance2_price_extractor(model_id, has_video_input=False),
             poll_interval=9,
-            max_poll_attempts=180,
         )
         return IO.NodeOutput(await download_url_to_video_output(response.content.video_url))
 
@@ -1228,12 +1428,27 @@ class ByteDance2FirstLastFrameNode(IO.ComfyNode):
                 IO.Image.Input(
                     "first_frame",
                     tooltip="First frame image for the video.",
+                    optional=True,
                 ),
                 IO.Image.Input(
                     "last_frame",
                     tooltip="Last frame image for the video.",
                     optional=True,
                 ),
+                IO.String.Input(
+                    "first_frame_asset_id",
+                    default="",
+                    tooltip="Seedance asset_id to use as the first frame. "
+                            "Mutually exclusive with the first_frame image input.",
+                    optional=True,
+                ),
+                IO.String.Input(
+                    "last_frame_asset_id",
+                    default="",
+                    tooltip="Seedance asset_id to use as the last frame. "
+                            "Mutually exclusive with the last_frame image input.",
+                    optional=True,
+                ),
                 IO.Int.Input(
                     "seed",
                     default=0,
@@ -1286,28 +1501,62 @@ class ByteDance2FirstLastFrameNode(IO.ComfyNode):
     async def execute(
         cls,
         model: dict,
-        first_frame: Input.Image,
         seed: int,
         watermark: bool,
+        first_frame: Input.Image | None = None,
         last_frame: Input.Image | None = None,
+        first_frame_asset_id: str = "",
+        last_frame_asset_id: str = "",
     ) -> IO.NodeOutput:
         validate_string(model["prompt"], strip_whitespace=True, min_length=1)
         model_id = SEEDANCE_MODELS[model["model"]]
 
+        first_frame_asset_id = first_frame_asset_id.strip()
+        last_frame_asset_id = last_frame_asset_id.strip()
+
+        if first_frame is not None and first_frame_asset_id:
+            raise ValueError("Provide only one of first_frame or first_frame_asset_id, not both.")
+        if first_frame is None and not first_frame_asset_id:
+            raise ValueError("Either first_frame or first_frame_asset_id is required.")
+        if last_frame is not None and last_frame_asset_id:
+            raise ValueError("Provide only one of last_frame or last_frame_asset_id, not both.")
+
+        asset_ids_to_resolve = [a for a in (first_frame_asset_id, last_frame_asset_id) if a]
+        image_assets: dict[str, str] = {}
+        if asset_ids_to_resolve:
+            image_assets, _, _ = await _resolve_reference_assets(cls, asset_ids_to_resolve)
+            for aid in asset_ids_to_resolve:
+                if aid not in image_assets:
+                    raise ValueError(f"Asset {aid} is not an Image asset.")
+
+        if first_frame_asset_id:
+            first_frame_url = image_assets[first_frame_asset_id]
+        else:
+            first_frame_url = await _seedance_virtual_library_upload_image_asset(
+                cls, first_frame, wait_label="Uploading first frame."
+            )
+
         content: list[TaskTextContent | TaskImageContent] = [
             TaskTextContent(text=model["prompt"]),
             TaskImageContent(
-                image_url=TaskImageContentUrl(
-                    url=await upload_image_to_comfyapi(cls, first_frame, wait_label="Uploading first frame.")
-                ),
+                image_url=TaskImageContentUrl(url=first_frame_url),
                 role="first_frame",
             ),
         ]
-        if last_frame is not None:
+        if last_frame_asset_id:
+            content.append(
+                TaskImageContent(
+                    image_url=TaskImageContentUrl(url=image_assets[last_frame_asset_id]),
+                    role="last_frame",
+                ),
+            )
+        elif last_frame is not None:
             content.append(
                 TaskImageContent(
                     image_url=TaskImageContentUrl(
-                        url=await upload_image_to_comfyapi(cls, last_frame, wait_label="Uploading last frame.")
+                        url=await _seedance_virtual_library_upload_image_asset(
+                            cls, last_frame, wait_label="Uploading last frame."
+                        )
                     ),
                     role="last_frame",
                 ),
@@ -1335,7 +1584,6 @@ class ByteDance2FirstLastFrameNode(IO.ComfyNode):
             status_extractor=lambda r: r.status,
             price_extractor=_seedance2_price_extractor(model_id, has_video_input=False),
             poll_interval=9,
-            max_poll_attempts=180,
         )
         return IO.NodeOutput(await download_url_to_video_output(response.content.video_url))
 
@@ -1385,6 +1633,24 @@ def _seedance2_reference_inputs(resolutions: list[str]):
             tooltip="Automatically downscale reference videos that exceed the model's pixel budget "
             "for the selected resolution. Aspect ratio is preserved; videos already within limits are untouched.",
         ),
+        IO.Autogrow.Input(
+            "reference_assets",
+            template=IO.Autogrow.TemplateNames(
+                IO.String.Input("reference_asset"),
+                names=[
+                    "asset_1",
+                    "asset_2",
+                    "asset_3",
+                    "asset_4",
+                    "asset_5",
+                    "asset_6",
+                    "asset_7",
+                    "asset_8",
+                    "asset_9",
+                ],
+                min=0,
+            ),
+        ),
     ]
 
 
@@ -1486,24 +1752,42 @@ class ByteDance2ReferenceNode(IO.ComfyNode):
         reference_images = model.get("reference_images", {})
         reference_videos = model.get("reference_videos", {})
         reference_audios = model.get("reference_audios", {})
+        reference_assets = model.get("reference_assets", {})
 
-        if not reference_images and not reference_videos:
-            raise ValueError("At least one reference image or video is required.")
+        reference_image_assets, reference_video_assets, reference_audio_assets = await _resolve_reference_assets(
+            cls, list(reference_assets.values())
+        )
+
+        if not reference_images and not reference_videos and not reference_image_assets and not reference_video_assets:
+            raise ValueError("At least one reference image or video or asset is required.")
+
+        total_images = len(reference_images) + len(reference_image_assets)
+        if total_images > 9:
+            raise ValueError(
+                f"Too many reference images: {total_images} "
+                f"(images={len(reference_images)}, image assets={len(reference_image_assets)}). Maximum is 9."
+            )
+        total_videos = len(reference_videos) + len(reference_video_assets)
+        if total_videos > 3:
+            raise ValueError(
+                f"Too many reference videos: {total_videos} "
+                f"(videos={len(reference_videos)}, video assets={len(reference_video_assets)}). Maximum is 3."
+            )
+        total_audios = len(reference_audios) + len(reference_audio_assets)
+        if total_audios > 3:
+            raise ValueError(
+                f"Too many reference audios: {total_audios} "
+                f"(audios={len(reference_audios)}, audio assets={len(reference_audio_assets)}). Maximum is 3."
+            )
 
         model_id = SEEDANCE_MODELS[model["model"]]
-        has_video_input = len(reference_videos) > 0
+        has_video_input = total_videos > 0
 
         if model.get("auto_downscale") and reference_videos:
-            max_px = (
-                SEEDANCE2_REF_VIDEO_PIXEL_LIMITS.get(model_id, {})
-                .get(model["resolution"], {})
-                .get("max")
-            )
+            max_px = SEEDANCE2_REF_VIDEO_PIXEL_LIMITS.get(model_id, {}).get(model["resolution"], {}).get("max")
             if max_px:
                 for key in reference_videos:
-                    reference_videos[key] = resize_video_to_pixel_budget(
-                        reference_videos[key], max_px
-                    )
+                    reference_videos[key] = resize_video_to_pixel_budget(reference_videos[key], max_px)
 
         total_video_duration = 0.0
         for i, key in enumerate(reference_videos, 1):
@@ -1531,16 +1815,27 @@ class ByteDance2ReferenceNode(IO.ComfyNode):
         if total_audio_duration > 15.1:
             raise ValueError(f"Total reference audio duration is {total_audio_duration:.1f}s. Maximum is 15.1 seconds.")
 
+        asset_labels = _build_asset_labels(
+            reference_assets,
+            reference_image_assets,
+            reference_video_assets,
+            reference_audio_assets,
+            len(reference_images),
+            len(reference_videos),
+            len(reference_audios),
+        )
+        prompt_text = _rewrite_asset_refs(model["prompt"], asset_labels)
+
         content: list[TaskTextContent | TaskImageContent | TaskVideoContent | TaskAudioContent] = [
-            TaskTextContent(text=model["prompt"]),
+            TaskTextContent(text=prompt_text),
         ]
         for i, key in enumerate(reference_images, 1):
             content.append(
                 TaskImageContent(
                     image_url=TaskImageContentUrl(
-                        url=await upload_image_to_comfyapi(
+                        url=await _seedance_virtual_library_upload_image_asset(
                             cls,
-                            image=reference_images[key],
+                            reference_images[key],
                             wait_label=f"Uploading image {i}",
                         ),
                     ),
@@ -1573,6 +1868,21 @@ class ByteDance2ReferenceNode(IO.ComfyNode):
                     ),
                 ),
             )
+        for url in reference_image_assets.values():
+            content.append(
+                TaskImageContent(
+                    image_url=TaskImageContentUrl(url=url),
+                    role="reference_image",
+                ),
+            )
+        for url in reference_video_assets.values():
+            content.append(
+                TaskVideoContent(video_url=TaskVideoContentUrl(url=url)),
+            )
+        for url in reference_audio_assets.values():
+            content.append(
+                TaskAudioContent(audio_url=TaskAudioContentUrl(url=url)),
+            )
         initial_response = await sync_op(
             cls,
             ApiEndpoint(path=BYTEPLUS_TASK_ENDPOINT, method="POST"),
@@ -1595,7 +1905,6 @@ class ByteDance2ReferenceNode(IO.ComfyNode):
             status_extractor=lambda r: r.status,
             price_extractor=_seedance2_price_extractor(model_id, has_video_input=has_video_input),
             poll_interval=9,
-            max_poll_attempts=180,
         )
         return IO.NodeOutput(await download_url_to_video_output(response.content.video_url))
 
@@ -1627,6 +1936,156 @@ async def process_video_task(
     return IO.NodeOutput(await download_url_to_video_output(response.content.video_url))
 
 
+class ByteDanceCreateImageAsset(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="ByteDanceCreateImageAsset",
+            display_name="ByteDance Create Image Asset",
+            category="api node/image/ByteDance",
+            description=(
+                "Create a Seedance 2.0 personal image asset. Uploads the input image and "
+                "registers it in the given asset group. If group_id is empty, runs a real-person "
+                "H5 authentication flow to create a new group before adding the asset."
+            ),
+            inputs=[
+                IO.Image.Input("image", tooltip="Image to register as a personal asset."),
+                IO.String.Input(
+                    "group_id",
+                    default="",
+                    tooltip="Reuse an existing Seedance asset group ID to skip repeated human verification for the "
+                    "same person. Leave empty to run real-person authentication in the browser and create a new group.",
+                ),
+                # IO.String.Input(
+                #     "name",
+                #     default="",
+                #     tooltip="Asset name (up to 64 characters).",
+                # ),
+            ],
+            outputs=[
+                IO.String.Output(display_name="asset_id"),
+                IO.String.Output(display_name="group_id"),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            # is_api_node=True,
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        image: Input.Image,
+        group_id: str = "",
+        # name: str = "",
+    ) -> IO.NodeOutput:
+        # if len(name) > 64:
+        #     raise ValueError("Name of asset can not be greater then 64 symbols")
+        validate_image_dimensions(image, min_width=300, max_width=6000, min_height=300, max_height=6000)
+        validate_image_aspect_ratio(image, min_ratio=(0.4, 1), max_ratio=(2.5, 1))
+        resolved_group = await _resolve_group_id(cls, group_id)
+        asset_id = await _create_seedance_asset(
+            cls,
+            group_id=resolved_group,
+            url=await upload_image_to_comfyapi(cls, image),
+            name="",
+            asset_type="Image",
+        )
+        await _wait_for_asset_active(cls, asset_id, resolved_group)
+        PromptServer.instance.send_progress_text(
+            f"Please save the asset_id and group_id for reuse.\n\nasset_id: {asset_id}\n\n"
+            f"group_id: {resolved_group}",
+            cls.hidden.unique_id,
+        )
+        return IO.NodeOutput(asset_id, resolved_group)
+
+
+class ByteDanceCreateVideoAsset(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="ByteDanceCreateVideoAsset",
+            display_name="ByteDance Create Video Asset",
+            category="api node/video/ByteDance",
+            description=(
+                "Create a Seedance 2.0 personal video asset. Uploads the input video and "
+                "registers it in the given asset group. If group_id is empty, runs a real-person "
+                "H5 authentication flow to create a new group before adding the asset."
+            ),
+            inputs=[
+                IO.Video.Input("video", tooltip="Video to register as a personal asset."),
+                IO.String.Input(
+                    "group_id",
+                    default="",
+                    tooltip="Reuse an existing Seedance asset group ID to skip repeated human verification for the "
+                    "same person. Leave empty to run real-person authentication in the browser and create a new group.",
+                ),
+                # IO.String.Input(
+                #     "name",
+                #     default="",
+                #     tooltip="Asset name (up to 64 characters).",
+                # ),
+            ],
+            outputs=[
+                IO.String.Output(display_name="asset_id"),
+                IO.String.Output(display_name="group_id"),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            # is_api_node=True,
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        video: Input.Video,
+        group_id: str = "",
+        # name: str = "",
+    ) -> IO.NodeOutput:
+        # if len(name) > 64:
+        #     raise ValueError("Name of asset can not be greater then 64 symbols")
+        validate_video_duration(video, min_duration=2, max_duration=15)
+        validate_video_dimensions(video, min_width=300, max_width=6000, min_height=300, max_height=6000)
+
+        w, h = video.get_dimensions()
+        if h > 0:
+            ratio = w / h
+            if not (0.4 <= ratio <= 2.5):
+                raise ValueError(f"Asset video aspect ratio (W/H) must be in [0.4, 2.5], got {ratio:.3f} ({w}x{h}).")
+        pixels = w * h
+        if not (409_600 <= pixels <= 927_408):
+            raise ValueError(
+                f"Asset video total pixels (W×H) must be in [409600, 927408], " f"got {pixels:,} ({w}x{h})."
+            )
+
+        fps = float(video.get_frame_rate())
+        if not (24 <= fps <= 60):
+            raise ValueError(f"Asset video FPS must be in [24, 60], got {fps:.2f}.")
+
+        resolved_group = await _resolve_group_id(cls, group_id)
+        asset_id = await _create_seedance_asset(
+            cls,
+            group_id=resolved_group,
+            url=await upload_video_to_comfyapi(cls, video),
+            name="",
+            asset_type="Video",
+        )
+        await _wait_for_asset_active(cls, asset_id, resolved_group)
+        PromptServer.instance.send_progress_text(
+            f"Please save the asset_id and group_id for reuse.\n\nasset_id: {asset_id}\n\n"
+            f"group_id: {resolved_group}",
+            cls.hidden.unique_id,
+        )
+        return IO.NodeOutput(asset_id, resolved_group)
+
+
 class ByteDanceExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[IO.ComfyNode]]:
@@ -1640,6 +2099,8 @@ class ByteDanceExtension(ComfyExtension):
             ByteDance2TextToVideoNode,
             ByteDance2FirstLastFrameNode,
             ByteDance2ReferenceNode,
+            ByteDanceCreateImageAsset,
+            ByteDanceCreateVideoAsset,
         ]
 
 

comfy_api_nodes/nodes_hitpaw.py

@@ -178,7 +178,6 @@ class HitPawGeneralImageEnhance(IO.ComfyNode):
             status_extractor=lambda x: x.data.status,
             price_extractor=lambda x: request_price,
             poll_interval=10.0,
-            max_poll_attempts=480,
         )
         return IO.NodeOutput(await download_url_to_image_tensor(final_response.data.res_url))
 
@@ -324,7 +323,6 @@ class HitPawVideoEnhance(IO.ComfyNode):
             status_extractor=lambda x: x.data.status,
             price_extractor=lambda x: request_price,
             poll_interval=10.0,
-            max_poll_attempts=320,
         )
         return IO.NodeOutput(await download_url_to_video_output(final_response.data.res_url))
 

comfy_api_nodes/nodes_kling.py

@@ -862,7 +862,7 @@ class OmniProTextToVideoNode(IO.ComfyNode):
                 ),
                 IO.Combo.Input("aspect_ratio", options=["16:9", "9:16", "1:1"]),
                 IO.Int.Input("duration", default=5, min=3, max=15, display_mode=IO.NumberDisplay.slider),
-                IO.Combo.Input("resolution", options=["1080p", "720p"], optional=True),
+                IO.Combo.Input("resolution", options=["4k", "1080p", "720p"], default="1080p", optional=True),
                 IO.DynamicCombo.Input(
                     "storyboards",
                     options=[
@@ -904,12 +904,13 @@ class OmniProTextToVideoNode(IO.ComfyNode):
                 depends_on=IO.PriceBadgeDepends(widgets=["duration", "resolution", "model_name", "generate_audio"]),
                 expr="""
                 (
-                  $mode := (widgets.resolution = "720p") ? "std" : "pro";
+                  $res := widgets.resolution;
+                  $mode := $res = "4k" ? "4k" : ($res = "720p" ? "std" : "pro");
                   $isV3 := $contains(widgets.model_name, "v3");
                   $audio := $isV3 and widgets.generate_audio;
                   $rates := $audio
-                    ? {"std": 0.112, "pro": 0.14}
-                    : {"std": 0.084, "pro": 0.112};
+                    ? {"std": 0.112, "pro": 0.14, "4k": 0.42}
+                    : {"std": 0.084, "pro": 0.112, "4k": 0.42};
                   {"type":"usd","usd": $lookup($rates, $mode) * widgets.duration}
                 )
                 """,
@@ -934,6 +935,8 @@ class OmniProTextToVideoNode(IO.ComfyNode):
                 raise ValueError("kling-video-o1 only supports durations of 5 or 10 seconds.")
             if generate_audio:
                 raise ValueError("kling-video-o1 does not support audio generation.")
+            if resolution == "4k":
+                raise ValueError("kling-video-o1 does not support 4k resolution.")
         stories_enabled = storyboards is not None and storyboards["storyboards"] != "disabled"
         if stories_enabled and model_name == "kling-video-o1":
             raise ValueError("kling-video-o1 does not support storyboards.")
@@ -963,6 +966,12 @@ class OmniProTextToVideoNode(IO.ComfyNode):
                     f"must equal the global duration ({duration}s)."
                 )
 
+        if resolution == "4k":
+            mode = "4k"
+        elif resolution == "1080p":
+            mode = "pro"
+        else:
+            mode = "std"
         response = await sync_op(
             cls,
             ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"),
@@ -972,7 +981,7 @@ class OmniProTextToVideoNode(IO.ComfyNode):
                 prompt=prompt,
                 aspect_ratio=aspect_ratio,
                 duration=str(duration),
-                mode="pro" if resolution == "1080p" else "std",
+                mode=mode,
                 multi_shot=multi_shot,
                 multi_prompt=multi_prompt_list,
                 shot_type="customize" if multi_shot else None,
@@ -1014,7 +1023,7 @@ class OmniProFirstLastFrameNode(IO.ComfyNode):
                     optional=True,
                     tooltip="Up to 6 additional reference images.",
                 ),
-                IO.Combo.Input("resolution", options=["1080p", "720p"], optional=True),
+                IO.Combo.Input("resolution", options=["4k", "1080p", "720p"], default="1080p", optional=True),
                 IO.DynamicCombo.Input(
                     "storyboards",
                     options=[
@@ -1061,12 +1070,13 @@ class OmniProFirstLastFrameNode(IO.ComfyNode):
                 depends_on=IO.PriceBadgeDepends(widgets=["duration", "resolution", "model_name", "generate_audio"]),
                 expr="""
                 (
-                  $mode := (widgets.resolution = "720p") ? "std" : "pro";
+                  $res := widgets.resolution;
+                  $mode := $res = "4k" ? "4k" : ($res = "720p" ? "std" : "pro");
                   $isV3 := $contains(widgets.model_name, "v3");
                   $audio := $isV3 and widgets.generate_audio;
                   $rates := $audio
-                    ? {"std": 0.112, "pro": 0.14}
-                    : {"std": 0.084, "pro": 0.112};
+                    ? {"std": 0.112, "pro": 0.14, "4k": 0.42}
+                    : {"std": 0.084, "pro": 0.112, "4k": 0.42};
                   {"type":"usd","usd": $lookup($rates, $mode) * widgets.duration}
                 )
                 """,
@@ -1093,6 +1103,8 @@ class OmniProFirstLastFrameNode(IO.ComfyNode):
                 raise ValueError("kling-video-o1 does not support durations greater than 10 seconds.")
             if generate_audio:
                 raise ValueError("kling-video-o1 does not support audio generation.")
+            if resolution == "4k":
+                raise ValueError("kling-video-o1 does not support 4k resolution.")
         stories_enabled = storyboards is not None and storyboards["storyboards"] != "disabled"
         if stories_enabled and model_name == "kling-video-o1":
             raise ValueError("kling-video-o1 does not support storyboards.")
@@ -1161,6 +1173,12 @@ class OmniProFirstLastFrameNode(IO.ComfyNode):
                 validate_image_aspect_ratio(i, (1, 2.5), (2.5, 1))
             for i in await upload_images_to_comfyapi(cls, reference_images, wait_label="Uploading reference frame(s)"):
                 image_list.append(OmniParamImage(image_url=i))
+        if resolution == "4k":
+            mode = "4k"
+        elif resolution == "1080p":
+            mode = "pro"
+        else:
+            mode = "std"
         response = await sync_op(
             cls,
             ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"),
@@ -1170,7 +1188,7 @@ class OmniProFirstLastFrameNode(IO.ComfyNode):
                 prompt=prompt,
                 duration=str(duration),
                 image_list=image_list,
-                mode="pro" if resolution == "1080p" else "std",
+                mode=mode,
                 sound="on" if generate_audio else "off",
                 multi_shot=multi_shot,
                 multi_prompt=multi_prompt_list,
@@ -1204,7 +1222,7 @@ class OmniProImageToVideoNode(IO.ComfyNode):
                     "reference_images",
                     tooltip="Up to 7 reference images.",
                 ),
-                IO.Combo.Input("resolution", options=["1080p", "720p"], optional=True),
+                IO.Combo.Input("resolution", options=["4k", "1080p", "720p"], default="1080p", optional=True),
                 IO.DynamicCombo.Input(
                     "storyboards",
                     options=[
@@ -1251,12 +1269,13 @@ class OmniProImageToVideoNode(IO.ComfyNode):
                 depends_on=IO.PriceBadgeDepends(widgets=["duration", "resolution", "model_name", "generate_audio"]),
                 expr="""
                 (
-                  $mode := (widgets.resolution = "720p") ? "std" : "pro";
+                  $res := widgets.resolution;
+                  $mode := $res = "4k" ? "4k" : ($res = "720p" ? "std" : "pro");
                   $isV3 := $contains(widgets.model_name, "v3");
                   $audio := $isV3 and widgets.generate_audio;
                   $rates := $audio
-                    ? {"std": 0.112, "pro": 0.14}
-                    : {"std": 0.084, "pro": 0.112};
+                    ? {"std": 0.112, "pro": 0.14, "4k": 0.42}
+                    : {"std": 0.084, "pro": 0.112, "4k": 0.42};
                   {"type":"usd","usd": $lookup($rates, $mode) * widgets.duration}
                 )
                 """,
@@ -1282,6 +1301,8 @@ class OmniProImageToVideoNode(IO.ComfyNode):
                 raise ValueError("kling-video-o1 does not support durations greater than 10 seconds.")
             if generate_audio:
                 raise ValueError("kling-video-o1 does not support audio generation.")
+            if resolution == "4k":
+                raise ValueError("kling-video-o1 does not support 4k resolution.")
         stories_enabled = storyboards is not None and storyboards["storyboards"] != "disabled"
         if stories_enabled and model_name == "kling-video-o1":
             raise ValueError("kling-video-o1 does not support storyboards.")
@@ -1320,6 +1341,12 @@ class OmniProImageToVideoNode(IO.ComfyNode):
         image_list: list[OmniParamImage] = []
         for i in await upload_images_to_comfyapi(cls, reference_images, wait_label="Uploading reference image"):
             image_list.append(OmniParamImage(image_url=i))
+        if resolution == "4k":
+            mode = "4k"
+        elif resolution == "1080p":
+            mode = "pro"
+        else:
+            mode = "std"
         response = await sync_op(
             cls,
             ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"),
@@ -1330,7 +1357,7 @@ class OmniProImageToVideoNode(IO.ComfyNode):
                 aspect_ratio=aspect_ratio,
                 duration=str(duration),
                 image_list=image_list,
-                mode="pro" if resolution == "1080p" else "std",
+                mode=mode,
                 sound="on" if generate_audio else "off",
                 multi_shot=multi_shot,
                 multi_prompt=multi_prompt_list,
@@ -2860,7 +2887,7 @@ class KlingVideoNode(IO.ComfyNode):
                         IO.DynamicCombo.Option(
                             "kling-v3",
                             [
-                                IO.Combo.Input("resolution", options=["1080p", "720p"]),
+                                IO.Combo.Input("resolution", options=["4k", "1080p", "720p"], default="1080p"),
                                 IO.Combo.Input(
                                     "aspect_ratio",
                                     options=["16:9", "9:16", "1:1"],
@@ -2913,7 +2940,11 @@ class KlingVideoNode(IO.ComfyNode):
                 ),
                 expr="""
                 (
-                  $rates := {"1080p": {"off": 0.112, "on": 0.168}, "720p": {"off": 0.084, "on": 0.126}};
+                  $rates := {
+                    "4k": {"off": 0.42, "on": 0.42},
+                    "1080p": {"off": 0.112, "on": 0.168},
+                    "720p": {"off": 0.084, "on": 0.126}
+                  };
                   $res := $lookup(widgets, "model.resolution");
                   $audio := widgets.generate_audio ? "on" : "off";
                   $rate := $lookup($lookup($rates, $res), $audio);
@@ -2943,7 +2974,12 @@ class KlingVideoNode(IO.ComfyNode):
         start_frame: Input.Image | None = None,
     ) -> IO.NodeOutput:
         _ = seed
-        mode = "pro" if model["resolution"] == "1080p" else "std"
+        if model["resolution"] == "4k":
+            mode = "4k"
+        elif model["resolution"] == "1080p":
+            mode = "pro"
+        else:
+            mode = "std"
         custom_multi_shot = False
         if multi_shot["multi_shot"] == "disabled":
             shot_type = None
@@ -3057,7 +3093,7 @@ class KlingFirstLastFrameNode(IO.ComfyNode):
                         IO.DynamicCombo.Option(
                             "kling-v3",
                             [
-                                IO.Combo.Input("resolution", options=["1080p", "720p"]),
+                                IO.Combo.Input("resolution", options=["4k", "1080p", "720p"], default="1080p"),
                             ],
                         ),
                     ],
@@ -3089,7 +3125,11 @@ class KlingFirstLastFrameNode(IO.ComfyNode):
                 ),
                 expr="""
                 (
-                  $rates := {"1080p": {"off": 0.112, "on": 0.168}, "720p": {"off": 0.084, "on": 0.126}};
+                  $rates := {
+                    "4k": {"off": 0.42, "on": 0.42},
+                    "1080p": {"off": 0.112, "on": 0.168},
+                    "720p": {"off": 0.084, "on": 0.126}
+                  };
                   $res := $lookup(widgets, "model.resolution");
                   $audio := widgets.generate_audio ? "on" : "off";
                   $rate := $lookup($lookup($rates, $res), $audio);
@@ -3118,6 +3158,12 @@ class KlingFirstLastFrameNode(IO.ComfyNode):
         validate_image_aspect_ratio(end_frame, (1, 2.5), (2.5, 1))
         image_url = await upload_image_to_comfyapi(cls, first_frame, wait_label="Uploading first frame")
         image_tail_url = await upload_image_to_comfyapi(cls, end_frame, wait_label="Uploading end frame")
+        if model["resolution"] == "4k":
+            mode = "4k"
+        elif model["resolution"] == "1080p":
+            mode = "pro"
+        else:
+            mode = "std"
         response = await sync_op(
             cls,
             ApiEndpoint(path="/proxy/kling/v1/videos/image2video", method="POST"),
@@ -3127,7 +3173,7 @@ class KlingFirstLastFrameNode(IO.ComfyNode):
                 image=image_url,
                 image_tail=image_tail_url,
                 prompt=prompt,
-                mode="pro" if model["resolution"] == "1080p" else "std",
+                mode=mode,
                 duration=str(duration),
                 sound="on" if generate_audio else "off",
             ),

comfy_api_nodes/nodes_magnific.py

@@ -230,7 +230,6 @@ class MagnificImageUpscalerCreativeNode(IO.ComfyNode):
             status_extractor=lambda x: x.status,
             price_extractor=lambda _: price_usd,
             poll_interval=10.0,
-            max_poll_attempts=480,
         )
         return IO.NodeOutput(await download_url_to_image_tensor(final_response.generated[0]))
 
@@ -391,7 +390,6 @@ class MagnificImageUpscalerPreciseV2Node(IO.ComfyNode):
             status_extractor=lambda x: x.status,
             price_extractor=lambda _: price_usd,
             poll_interval=10.0,
-            max_poll_attempts=480,
         )
         return IO.NodeOutput(await download_url_to_image_tensor(final_response.generated[0]))
 
@@ -541,7 +539,6 @@ class MagnificImageStyleTransferNode(IO.ComfyNode):
             response_model=TaskResponse,
             status_extractor=lambda x: x.status,
             poll_interval=10.0,
-            max_poll_attempts=480,
         )
         return IO.NodeOutput(await download_url_to_image_tensor(final_response.generated[0]))
 
@@ -782,7 +779,6 @@ class MagnificImageRelightNode(IO.ComfyNode):
             response_model=TaskResponse,
             status_extractor=lambda x: x.status,
             poll_interval=10.0,
-            max_poll_attempts=480,
         )
         return IO.NodeOutput(await download_url_to_image_tensor(final_response.generated[0]))
 
@@ -924,7 +920,6 @@ class MagnificImageSkinEnhancerNode(IO.ComfyNode):
             response_model=TaskResponse,
             status_extractor=lambda x: x.status,
             poll_interval=10.0,
-            max_poll_attempts=480,
         )
         return IO.NodeOutput(await download_url_to_image_tensor(final_response.generated[0]))
 

comfy_api_nodes/nodes_moonvalley.py

@@ -1,534 +0,0 @@
-import logging
-
-from typing_extensions import override
-
-from comfy_api.latest import IO, ComfyExtension, Input
-from comfy_api_nodes.apis.moonvalley import (
-    MoonvalleyPromptResponse,
-    MoonvalleyTextToVideoInferenceParams,
-    MoonvalleyTextToVideoRequest,
-    MoonvalleyVideoToVideoInferenceParams,
-    MoonvalleyVideoToVideoRequest,
-)
-from comfy_api_nodes.util import (
-    ApiEndpoint,
-    download_url_to_video_output,
-    poll_op,
-    sync_op,
-    trim_video,
-    upload_images_to_comfyapi,
-    upload_video_to_comfyapi,
-    validate_container_format_is_mp4,
-    validate_image_dimensions,
-    validate_string,
-)
-
-API_UPLOADS_ENDPOINT = "/proxy/moonvalley/uploads"
-API_PROMPTS_ENDPOINT = "/proxy/moonvalley/prompts"
-API_VIDEO2VIDEO_ENDPOINT = "/proxy/moonvalley/prompts/video-to-video"
-API_TXT2VIDEO_ENDPOINT = "/proxy/moonvalley/prompts/text-to-video"
-API_IMG2VIDEO_ENDPOINT = "/proxy/moonvalley/prompts/image-to-video"
-
-MIN_WIDTH = 300
-MIN_HEIGHT = 300
-
-MAX_WIDTH = 10000
-MAX_HEIGHT = 10000
-
-MIN_VID_WIDTH = 300
-MIN_VID_HEIGHT = 300
-
-MAX_VID_WIDTH = 10000
-MAX_VID_HEIGHT = 10000
-
-MAX_VIDEO_SIZE = 1024 * 1024 * 1024  # 1 GB max for in-memory video processing
-
-MOONVALLEY_MAREY_MAX_PROMPT_LENGTH = 5000
-
-
-def is_valid_task_creation_response(response: MoonvalleyPromptResponse) -> bool:
-    """Verifies that the initial response contains a task ID."""
-    return bool(response.id)
-
-
-def validate_task_creation_response(response) -> None:
-    if not is_valid_task_creation_response(response):
-        error_msg = f"Moonvalley Marey API: Initial request failed. Code: {response.code}, Message: {response.message}, Data: {response}"
-        logging.error(error_msg)
-        raise RuntimeError(error_msg)
-
-
-def validate_video_to_video_input(video: Input.Video) -> Input.Video:
-    """
-    Validates and processes video input for Moonvalley Video-to-Video generation.
-
-    Args:
-        video: Input video to validate
-
-    Returns:
-        Validated and potentially trimmed video
-
-    Raises:
-        ValueError: If video doesn't meet requirements
-        MoonvalleyApiError: If video duration is too short
-    """
-    width, height = _get_video_dimensions(video)
-    _validate_video_dimensions(width, height)
-    validate_container_format_is_mp4(video)
-
-    return _validate_and_trim_duration(video)
-
-
-def _get_video_dimensions(video: Input.Video) -> tuple[int, int]:
-    """Extracts video dimensions with error handling."""
-    try:
-        return video.get_dimensions()
-    except Exception as e:
-        logging.error("Error getting dimensions of video: %s", e)
-        raise ValueError(f"Cannot get video dimensions: {e}") from e
-
-
-def _validate_video_dimensions(width: int, height: int) -> None:
-    """Validates video dimensions meet Moonvalley V2V requirements."""
-    supported_resolutions = {
-        (1920, 1080),
-        (1080, 1920),
-        (1152, 1152),
-        (1536, 1152),
-        (1152, 1536),
-    }
-
-    if (width, height) not in supported_resolutions:
-        supported_list = ", ".join([f"{w}x{h}" for w, h in sorted(supported_resolutions)])
-        raise ValueError(f"Resolution {width}x{height} not supported. Supported: {supported_list}")
-
-
-def _validate_and_trim_duration(video: Input.Video) -> Input.Video:
-    """Validates video duration and trims to 5 seconds if needed."""
-    duration = video.get_duration()
-    _validate_minimum_duration(duration)
-    return _trim_if_too_long(video, duration)
-
-
-def _validate_minimum_duration(duration: float) -> None:
-    """Ensures video is at least 5 seconds long."""
-    if duration < 5:
-        raise ValueError("Input video must be at least 5 seconds long.")
-
-
-def _trim_if_too_long(video: Input.Video, duration: float) -> Input.Video:
-    """Trims video to 5 seconds if longer."""
-    if duration > 5:
-        return trim_video(video, 5)
-    return video
-
-
-def parse_width_height_from_res(resolution: str):
-    # Accepts a string like "16:9 (1920 x 1080)" and returns width, height as a dict
-    res_map = {
-        "16:9 (1920 x 1080)": {"width": 1920, "height": 1080},
-        "9:16 (1080 x 1920)": {"width": 1080, "height": 1920},
-        "1:1 (1152 x 1152)": {"width": 1152, "height": 1152},
-        "4:3 (1536 x 1152)": {"width": 1536, "height": 1152},
-        "3:4 (1152 x 1536)": {"width": 1152, "height": 1536},
-        # "21:9 (2560 x 1080)": {"width": 2560, "height": 1080},
-    }
-    return res_map.get(resolution, {"width": 1920, "height": 1080})
-
-
-def parse_control_parameter(value):
-    control_map = {
-        "Motion Transfer": "motion_control",
-        "Canny": "canny_control",
-        "Pose Transfer": "pose_control",
-        "Depth": "depth_control",
-    }
-    return control_map.get(value, control_map["Motion Transfer"])
-
-
-async def get_response(cls: type[IO.ComfyNode], task_id: str) -> MoonvalleyPromptResponse:
-    return await poll_op(
-        cls,
-        ApiEndpoint(path=f"{API_PROMPTS_ENDPOINT}/{task_id}"),
-        response_model=MoonvalleyPromptResponse,
-        status_extractor=lambda r: (r.status if r and r.status else None),
-        poll_interval=16.0,
-        max_poll_attempts=240,
-    )
-
-
-class MoonvalleyImg2VideoNode(IO.ComfyNode):
-
-    @classmethod
-    def define_schema(cls) -> IO.Schema:
-        return IO.Schema(
-            node_id="MoonvalleyImg2VideoNode",
-            display_name="Moonvalley Marey Image to Video",
-            category="api node/video/Moonvalley Marey",
-            description="Moonvalley Marey Image to Video Node",
-            inputs=[
-                IO.Image.Input(
-                    "image",
-                    tooltip="The reference image used to generate the video",
-                ),
-                IO.String.Input(
-                    "prompt",
-                    multiline=True,
-                ),
-                IO.String.Input(
-                    "negative_prompt",
-                    multiline=True,
-                    default="<synthetic> <scene cut> gopro, bright, contrast, static, overexposed, vignette, "
-                    "artifacts, still, noise, texture, scanlines, videogame, 360 camera, VR, transition, "
-                    "flare, saturation, distorted, warped, wide angle, saturated, vibrant, glowing, "
-                    "cross dissolve, cheesy, ugly hands, mutated hands, mutant, disfigured, extra fingers, "
-                    "blown out, horrible, blurry, worst quality, bad, dissolve, melt, fade in, fade out, "
-                    "wobbly, weird, low quality, plastic, stock footage, video camera, boring",
-                    tooltip="Negative prompt text",
-                ),
-                IO.Combo.Input(
-                    "resolution",
-                    options=[
-                        "16:9 (1920 x 1080)",
-                        "9:16 (1080 x 1920)",
-                        "1:1 (1152 x 1152)",
-                        "4:3 (1536 x 1152)",
-                        "3:4 (1152 x 1536)",
-                        # "21:9 (2560 x 1080)",
-                    ],
-                    default="16:9 (1920 x 1080)",
-                    tooltip="Resolution of the output video",
-                ),
-                IO.Float.Input(
-                    "prompt_adherence",
-                    default=4.5,
-                    min=1.0,
-                    max=20.0,
-                    step=1.0,
-                    tooltip="Guidance scale for generation control",
-                ),
-                IO.Int.Input(
-                    "seed",
-                    default=9,
-                    min=0,
-                    max=4294967295,
-                    step=1,
-                    display_mode=IO.NumberDisplay.number,
-                    tooltip="Random seed value",
-                    control_after_generate=True,
-                ),
-                IO.Int.Input(
-                    "steps",
-                    default=80,
-                    min=75,  # steps should be greater or equal to cooldown_steps(75) + warmup_steps(0)
-                    max=100,
-                    step=1,
-                    tooltip="Number of denoising steps",
-                ),
-            ],
-            outputs=[IO.Video.Output()],
-            hidden=[
-                IO.Hidden.auth_token_comfy_org,
-                IO.Hidden.api_key_comfy_org,
-                IO.Hidden.unique_id,
-            ],
-            is_api_node=True,
-            price_badge=IO.PriceBadge(
-                depends_on=IO.PriceBadgeDepends(),
-                expr="""{"type":"usd","usd": 1.5}""",
-            ),
-        )
-
-    @classmethod
-    async def execute(
-        cls,
-        image: Input.Image,
-        prompt: str,
-        negative_prompt: str,
-        resolution: str,
-        prompt_adherence: float,
-        seed: int,
-        steps: int,
-    ) -> IO.NodeOutput:
-        validate_image_dimensions(image, min_width=300, min_height=300, max_height=MAX_HEIGHT, max_width=MAX_WIDTH)
-        validate_string(prompt, min_length=1, max_length=MOONVALLEY_MAREY_MAX_PROMPT_LENGTH)
-        validate_string(negative_prompt, field_name="negative_prompt", max_length=MOONVALLEY_MAREY_MAX_PROMPT_LENGTH)
-        width_height = parse_width_height_from_res(resolution)
-
-        inference_params = MoonvalleyTextToVideoInferenceParams(
-            negative_prompt=negative_prompt,
-            steps=steps,
-            seed=seed,
-            guidance_scale=prompt_adherence,
-            width=width_height["width"],
-            height=width_height["height"],
-            use_negative_prompts=True,
-        )
-
-        # Get MIME type from tensor - assuming PNG format for image tensors
-        mime_type = "image/png"
-        image_url = (await upload_images_to_comfyapi(cls, image, max_images=1, mime_type=mime_type))[0]
-        task_creation_response = await sync_op(
-            cls,
-            endpoint=ApiEndpoint(path=API_IMG2VIDEO_ENDPOINT, method="POST"),
-            response_model=MoonvalleyPromptResponse,
-            data=MoonvalleyTextToVideoRequest(
-                image_url=image_url, prompt_text=prompt, inference_params=inference_params
-            ),
-        )
-        validate_task_creation_response(task_creation_response)
-        final_response = await get_response(cls, task_creation_response.id)
-        video = await download_url_to_video_output(final_response.output_url)
-        return IO.NodeOutput(video)
-
-
-class MoonvalleyVideo2VideoNode(IO.ComfyNode):
-
-    @classmethod
-    def define_schema(cls) -> IO.Schema:
-        return IO.Schema(
-            node_id="MoonvalleyVideo2VideoNode",
-            display_name="Moonvalley Marey Video to Video",
-            category="api node/video/Moonvalley Marey",
-            description="",
-            inputs=[
-                IO.String.Input(
-                    "prompt",
-                    multiline=True,
-                    tooltip="Describes the video to generate",
-                ),
-                IO.String.Input(
-                    "negative_prompt",
-                    multiline=True,
-                    default="<synthetic> <scene cut> gopro, bright, contrast, static, overexposed, vignette, "
-                    "artifacts, still, noise, texture, scanlines, videogame, 360 camera, VR, transition, "
-                    "flare, saturation, distorted, warped, wide angle, saturated, vibrant, glowing, "
-                    "cross dissolve, cheesy, ugly hands, mutated hands, mutant, disfigured, extra fingers, "
-                    "blown out, horrible, blurry, worst quality, bad, dissolve, melt, fade in, fade out, "
-                    "wobbly, weird, low quality, plastic, stock footage, video camera, boring",
-                    tooltip="Negative prompt text",
-                ),
-                IO.Int.Input(
-                    "seed",
-                    default=9,
-                    min=0,
-                    max=4294967295,
-                    step=1,
-                    display_mode=IO.NumberDisplay.number,
-                    tooltip="Random seed value",
-                    control_after_generate=False,
-                ),
-                IO.Video.Input(
-                    "video",
-                    tooltip="The reference video used to generate the output video. Must be at least 5 seconds long. "
-                    "Videos longer than 5s will be automatically trimmed. Only MP4 format supported.",
-                ),
-                IO.Combo.Input(
-                    "control_type",
-                    options=["Motion Transfer", "Pose Transfer"],
-                    default="Motion Transfer",
-                    optional=True,
-                ),
-                IO.Int.Input(
-                    "motion_intensity",
-                    default=100,
-                    min=0,
-                    max=100,
-                    step=1,
-                    tooltip="Only used if control_type is 'Motion Transfer'",
-                    optional=True,
-                ),
-                IO.Int.Input(
-                    "steps",
-                    default=60,
-                    min=60,  # steps should be greater or equal to cooldown_steps(36) + warmup_steps(24)
-                    max=100,
-                    step=1,
-                    display_mode=IO.NumberDisplay.number,
-                    tooltip="Number of inference steps",
-                ),
-            ],
-            outputs=[IO.Video.Output()],
-            hidden=[
-                IO.Hidden.auth_token_comfy_org,
-                IO.Hidden.api_key_comfy_org,
-                IO.Hidden.unique_id,
-            ],
-            is_api_node=True,
-            price_badge=IO.PriceBadge(
-                depends_on=IO.PriceBadgeDepends(),
-                expr="""{"type":"usd","usd": 2.25}""",
-            ),
-        )
-
-    @classmethod
-    async def execute(
-        cls,
-        prompt: str,
-        negative_prompt: str,
-        seed: int,
-        video: Input.Video | None = None,
-        control_type: str = "Motion Transfer",
-        motion_intensity: int | None = 100,
-        steps=60,
-        prompt_adherence=4.5,
-    ) -> IO.NodeOutput:
-        validated_video = validate_video_to_video_input(video)
-        video_url = await upload_video_to_comfyapi(cls, validated_video)
-        validate_string(prompt, min_length=1, max_length=MOONVALLEY_MAREY_MAX_PROMPT_LENGTH)
-        validate_string(negative_prompt, field_name="negative_prompt", max_length=MOONVALLEY_MAREY_MAX_PROMPT_LENGTH)
-
-        # Only include motion_intensity for Motion Transfer
-        control_params = {}
-        if control_type == "Motion Transfer" and motion_intensity is not None:
-            control_params["motion_intensity"] = motion_intensity
-
-        inference_params = MoonvalleyVideoToVideoInferenceParams(
-            negative_prompt=negative_prompt,
-            seed=seed,
-            control_params=control_params,
-            steps=steps,
-            guidance_scale=prompt_adherence,
-        )
-
-        task_creation_response = await sync_op(
-            cls,
-            endpoint=ApiEndpoint(path=API_VIDEO2VIDEO_ENDPOINT, method="POST"),
-            response_model=MoonvalleyPromptResponse,
-            data=MoonvalleyVideoToVideoRequest(
-                control_type=parse_control_parameter(control_type),
-                video_url=video_url,
-                prompt_text=prompt,
-                inference_params=inference_params,
-            ),
-        )
-        validate_task_creation_response(task_creation_response)
-        final_response = await get_response(cls, task_creation_response.id)
-        return IO.NodeOutput(await download_url_to_video_output(final_response.output_url))
-
-
-class MoonvalleyTxt2VideoNode(IO.ComfyNode):
-
-    @classmethod
-    def define_schema(cls) -> IO.Schema:
-        return IO.Schema(
-            node_id="MoonvalleyTxt2VideoNode",
-            display_name="Moonvalley Marey Text to Video",
-            category="api node/video/Moonvalley Marey",
-            description="",
-            inputs=[
-                IO.String.Input(
-                    "prompt",
-                    multiline=True,
-                ),
-                IO.String.Input(
-                    "negative_prompt",
-                    multiline=True,
-                    default="<synthetic> <scene cut> gopro, bright, contrast, static, overexposed, vignette, "
-                    "artifacts, still, noise, texture, scanlines, videogame, 360 camera, VR, transition, "
-                    "flare, saturation, distorted, warped, wide angle, saturated, vibrant, glowing, "
-                    "cross dissolve, cheesy, ugly hands, mutated hands, mutant, disfigured, extra fingers, "
-                    "blown out, horrible, blurry, worst quality, bad, dissolve, melt, fade in, fade out, "
-                    "wobbly, weird, low quality, plastic, stock footage, video camera, boring",
-                    tooltip="Negative prompt text",
-                ),
-                IO.Combo.Input(
-                    "resolution",
-                    options=[
-                        "16:9 (1920 x 1080)",
-                        "9:16 (1080 x 1920)",
-                        "1:1 (1152 x 1152)",
-                        "4:3 (1536 x 1152)",
-                        "3:4 (1152 x 1536)",
-                        "21:9 (2560 x 1080)",
-                    ],
-                    default="16:9 (1920 x 1080)",
-                    tooltip="Resolution of the output video",
-                ),
-                IO.Float.Input(
-                    "prompt_adherence",
-                    default=4.0,
-                    min=1.0,
-                    max=20.0,
-                    step=1.0,
-                    tooltip="Guidance scale for generation control",
-                ),
-                IO.Int.Input(
-                    "seed",
-                    default=9,
-                    min=0,
-                    max=4294967295,
-                    step=1,
-                    display_mode=IO.NumberDisplay.number,
-                    control_after_generate=True,
-                    tooltip="Random seed value",
-                ),
-                IO.Int.Input(
-                    "steps",
-                    default=80,
-                    min=75,  # steps should be greater or equal to cooldown_steps(75) + warmup_steps(0)
-                    max=100,
-                    step=1,
-                    tooltip="Inference steps",
-                ),
-            ],
-            outputs=[IO.Video.Output()],
-            hidden=[
-                IO.Hidden.auth_token_comfy_org,
-                IO.Hidden.api_key_comfy_org,
-                IO.Hidden.unique_id,
-            ],
-            is_api_node=True,
-            price_badge=IO.PriceBadge(
-                depends_on=IO.PriceBadgeDepends(),
-                expr="""{"type":"usd","usd": 1.5}""",
-            ),
-        )
-
-    @classmethod
-    async def execute(
-        cls,
-        prompt: str,
-        negative_prompt: str,
-        resolution: str,
-        prompt_adherence: float,
-        seed: int,
-        steps: int,
-    ) -> IO.NodeOutput:
-        validate_string(prompt, min_length=1, max_length=MOONVALLEY_MAREY_MAX_PROMPT_LENGTH)
-        validate_string(negative_prompt, field_name="negative_prompt", max_length=MOONVALLEY_MAREY_MAX_PROMPT_LENGTH)
-        width_height = parse_width_height_from_res(resolution)
-
-        inference_params = MoonvalleyTextToVideoInferenceParams(
-            negative_prompt=negative_prompt,
-            steps=steps,
-            seed=seed,
-            guidance_scale=prompt_adherence,
-            num_frames=128,
-            width=width_height["width"],
-            height=width_height["height"],
-        )
-
-        task_creation_response = await sync_op(
-            cls,
-            endpoint=ApiEndpoint(path=API_TXT2VIDEO_ENDPOINT, method="POST"),
-            response_model=MoonvalleyPromptResponse,
-            data=MoonvalleyTextToVideoRequest(prompt_text=prompt, inference_params=inference_params),
-        )
-        validate_task_creation_response(task_creation_response)
-        final_response = await get_response(cls, task_creation_response.id)
-        return IO.NodeOutput(await download_url_to_video_output(final_response.output_url))
-
-
-class MoonvalleyExtension(ComfyExtension):
-    @override
-    async def get_node_list(self) -> list[type[IO.ComfyNode]]:
-        return [
-            MoonvalleyImg2VideoNode,
-            MoonvalleyTxt2VideoNode,
-            MoonvalleyVideo2VideoNode,
-        ]
-
-
-async def comfy_entrypoint() -> MoonvalleyExtension:
-    return MoonvalleyExtension()

comfy_api_nodes/nodes_openai.py

@@ -357,13 +357,17 @@ def calculate_tokens_price_image_1_5(response: OpenAIImageGenerationResponse) ->
     return ((response.usage.input_tokens * 8.0) + (response.usage.output_tokens * 32.0)) / 1_000_000.0
 
 
+def calculate_tokens_price_image_2_0(response: OpenAIImageGenerationResponse) -> float | None:
+    return ((response.usage.input_tokens * 8.0) + (response.usage.output_tokens * 30.0)) / 1_000_000.0
+
+
 class OpenAIGPTImage1(IO.ComfyNode):
 
     @classmethod
     def define_schema(cls):
         return IO.Schema(
             node_id="OpenAIGPTImage1",
-            display_name="OpenAI GPT Image 1.5",
+            display_name="OpenAI GPT Image 2",
             category="api node/image/OpenAI",
             description="Generates images synchronously via OpenAI's GPT Image endpoint.",
             inputs=[
@@ -401,8 +405,19 @@ class OpenAIGPTImage1(IO.ComfyNode):
                 IO.Combo.Input(
                     "size",
                     default="auto",
-                    options=["auto", "1024x1024", "1024x1536", "1536x1024"],
-                    tooltip="Image size",
+                    options=[
+                        "auto",
+                        "1024x1024",
+                        "1024x1536",
+                        "1536x1024",
+                        "2048x2048",
+                        "2048x1152",
+                        "1152x2048",
+                        "3840x2160",
+                        "2160x3840",
+                        "Custom",
+                    ],
+                    tooltip="Image size. Select 'Custom' to use the custom width and height (GPT Image 2 only).",
                     optional=True,
                 ),
                 IO.Int.Input(
@@ -427,8 +442,26 @@ class OpenAIGPTImage1(IO.ComfyNode):
                 ),
                 IO.Combo.Input(
                     "model",
-                    options=["gpt-image-1", "gpt-image-1.5"],
-                    default="gpt-image-1.5",
+                    options=["gpt-image-1", "gpt-image-1.5", "gpt-image-2"],
+                    default="gpt-image-2",
+                    optional=True,
+                ),
+                IO.Int.Input(
+                    "custom_width",
+                    default=1024,
+                    min=1024,
+                    max=3840,
+                    step=16,
+                    tooltip="Used only when `size` is 'Custom'. Must be a multiple of 16 (GPT Image 2 only).",
+                    optional=True,
+                ),
+                IO.Int.Input(
+                    "custom_height",
+                    default=1024,
+                    min=1024,
+                    max=3840,
+                    step=16,
+                    tooltip="Used only when `size` is 'Custom'. Must be a multiple of 16 (GPT Image 2 only).",
                     optional=True,
                 ),
             ],
@@ -442,23 +475,36 @@ class OpenAIGPTImage1(IO.ComfyNode):
             ],
             is_api_node=True,
             price_badge=IO.PriceBadge(
-                depends_on=IO.PriceBadgeDepends(widgets=["quality", "n"]),
+                depends_on=IO.PriceBadgeDepends(widgets=["quality", "n", "model"]),
                 expr="""
                 (
                   $ranges := {
-                    "low":    [0.011, 0.02],
-                    "medium": [0.046, 0.07],
-                    "high":   [0.167, 0.3]
+                    "gpt-image-1": {
+                      "low":    [0.011, 0.02],
+                      "medium": [0.042, 0.07],
+                      "high":   [0.167, 0.25]
+                    },
+                    "gpt-image-1.5": {
+                      "low":    [0.009, 0.02],
+                      "medium": [0.034, 0.062],
+                      "high":   [0.133, 0.22]
+                    },
+                    "gpt-image-2": {
+                      "low":    [0.0048, 0.019],
+                      "medium": [0.041, 0.168],
+                      "high":   [0.165, 0.67]
+                    }
                   };
-                  $range := $lookup($ranges, widgets.quality);
-                  $n := widgets.n;
+                  $range := $lookup($lookup($ranges, widgets.model), widgets.quality);
+                  $nRaw := widgets.n;
+                  $n := ($nRaw != null and $nRaw != 0) ? $nRaw : 1;
                   ($n = 1)
-                    ? {"type":"range_usd","min_usd": $range[0], "max_usd": $range[1]}
+                    ? {"type":"range_usd","min_usd": $range[0], "max_usd": $range[1], "format": {"approximate": true}}
                     : {
                         "type":"range_usd",
-                        "min_usd": $range[0],
-                        "max_usd": $range[1],
-                        "format": { "suffix": " x " & $string($n) & "/Run" }
+                        "min_usd": $range[0] * $n,
+                        "max_usd": $range[1] * $n,
+                        "format": { "suffix": "/Run", "approximate": true }
                       }
                 )
                 """,
@@ -476,6 +522,8 @@ class OpenAIGPTImage1(IO.ComfyNode):
         mask: Input.Image | None = None,
         n: int = 1,
         size: str = "1024x1024",
+        custom_width: int = 1024,
+        custom_height: int = 1024,
         model: str = "gpt-image-1",
     ) -> IO.NodeOutput:
         validate_string(prompt, strip_whitespace=False)
@@ -483,10 +531,36 @@ class OpenAIGPTImage1(IO.ComfyNode):
         if mask is not None and image is None:
             raise ValueError("Cannot use a mask without an input image")
 
+        if size == "Custom":
+            if model != "gpt-image-2":
+                raise ValueError("Custom resolution is only supported by GPT Image 2 model")
+            if custom_width % 16 != 0 or custom_height % 16 != 0:
+                raise ValueError(f"Custom width and height must be multiples of 16, got {custom_width}x{custom_height}")
+            if max(custom_width, custom_height) > 3840:
+                raise ValueError(f"Custom resolution max edge must be <= 3840, got {custom_width}x{custom_height}")
+            ratio = max(custom_width, custom_height) / min(custom_width, custom_height)
+            if ratio > 3:
+                raise ValueError(
+                    f"Custom resolution aspect ratio must not exceed 3:1, got {custom_width}x{custom_height}"
+                )
+            total_pixels = custom_width * custom_height
+            if not 655_360 <= total_pixels <= 8_294_400:
+                raise ValueError(
+                    f"Custom resolution total pixels must be between 655,360 and 8,294,400, got {total_pixels}"
+                )
+            size = f"{custom_width}x{custom_height}"
+        elif model in ("gpt-image-1", "gpt-image-1.5"):
+            if size not in ("auto", "1024x1024", "1024x1536", "1536x1024"):
+                raise ValueError(f"Resolution {size} is only supported by GPT Image 2 model")
+
         if model == "gpt-image-1":
             price_extractor = calculate_tokens_price_image_1
         elif model == "gpt-image-1.5":
             price_extractor = calculate_tokens_price_image_1_5
+        elif model == "gpt-image-2":
+            price_extractor = calculate_tokens_price_image_2_0
+            if background == "transparent":
+                raise ValueError("Transparent background is not supported for GPT Image 2 model")
         else:
             raise ValueError(f"Unknown model: {model}")
 

comfy_api_nodes/nodes_sora.py

@@ -33,9 +33,13 @@ class OpenAIVideoSora2(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="OpenAIVideoSora2",
-            display_name="OpenAI Sora - Video",
+            display_name="OpenAI Sora - Video (Deprecated)",
             category="api node/video/Sora",
-            description="OpenAI video and audio generation.",
+            description=(
+                "OpenAI video and audio generation.\n\n"
+                "DEPRECATION NOTICE: OpenAI will stop serving the Sora v2 API in September 2026. "
+                "This node will be removed from ComfyUI at that time."
+            ),
             inputs=[
                 IO.Combo.Input(
                     "model",

comfy_api_nodes/nodes_topaz.py

@@ -36,11 +36,15 @@ from comfy_api_nodes.util import (
 )
 
 UPSCALER_MODELS_MAP = {
+    "Astra 2": "ast-2",
     "Starlight (Astra) Fast": "slf-1",
     "Starlight (Astra) Creative": "slc-1",
     "Starlight Precise 2.5": "slp-2.5",
 }
 
+AST2_MAX_FRAMES = 9000
+AST2_MAX_FRAMES_WITH_PROMPT = 450
+
 
 class TopazImageEnhance(IO.ComfyNode):
     @classmethod
@@ -230,13 +234,20 @@ class TopazVideoEnhance(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="TopazVideoEnhance",
-            display_name="Topaz Video Enhance",
+            display_name="Topaz Video Enhance (Legacy)",
             category="api node/video/Topaz",
             description="Breathe new life into video with powerful upscaling and recovery technology.",
             inputs=[
                 IO.Video.Input("video"),
                 IO.Boolean.Input("upscaler_enabled", default=True),
-                IO.Combo.Input("upscaler_model", options=list(UPSCALER_MODELS_MAP.keys())),
+                IO.Combo.Input(
+                    "upscaler_model",
+                    options=[
+                        "Starlight (Astra) Fast",
+                        "Starlight (Astra) Creative",
+                        "Starlight Precise 2.5",
+                    ],
+                ),
                 IO.Combo.Input("upscaler_resolution", options=["FullHD (1080p)", "4K (2160p)"]),
                 IO.Combo.Input(
                     "upscaler_creativity",
@@ -304,6 +315,7 @@ class TopazVideoEnhance(IO.ComfyNode):
                 IO.Hidden.unique_id,
             ],
             is_api_node=True,
+            is_deprecated=True,
         )
 
     @classmethod
@@ -453,7 +465,350 @@ class TopazVideoEnhance(IO.ComfyNode):
             progress_extractor=lambda x: getattr(x, "progress", 0),
             price_extractor=lambda x: (x.estimates.cost[0] * 0.08 if x.estimates and x.estimates.cost[0] else None),
             poll_interval=10.0,
-            max_poll_attempts=320,
+        )
+        return IO.NodeOutput(await download_url_to_video_output(final_response.download.url))
+
+
+class TopazVideoEnhanceV2(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="TopazVideoEnhanceV2",
+            display_name="Topaz Video Enhance",
+            category="api node/video/Topaz",
+            description="Breathe new life into video with powerful upscaling and recovery technology.",
+            inputs=[
+                IO.Video.Input("video"),
+                IO.DynamicCombo.Input(
+                    "upscaler_model",
+                    options=[
+                        IO.DynamicCombo.Option(
+                            "Astra 2",
+                            [
+                                IO.Combo.Input("upscaler_resolution", options=["FullHD (1080p)", "4K (2160p)"]),
+                                IO.Float.Input(
+                                    "creativity",
+                                    default=0.5,
+                                    min=0.0,
+                                    max=1.0,
+                                    step=0.1,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Creative strength of the upscale.",
+                                ),
+                                IO.String.Input(
+                                    "prompt",
+                                    multiline=True,
+                                    default="",
+                                    tooltip="Optional descriptive (not instructive) scene prompt."
+                                    f"Capping input at {AST2_MAX_FRAMES_WITH_PROMPT} frames (~15s @ 30fps) when set.",
+                                ),
+                                IO.Float.Input(
+                                    "sharp",
+                                    default=0.5,
+                                    min=0.0,
+                                    max=1.0,
+                                    step=0.01,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Pre-enhance sharpness: "
+                                    "0.0=Gaussian blur, 0.5=passthrough (default), 1.0=USM sharpening.",
+                                    advanced=True,
+                                ),
+                                IO.Float.Input(
+                                    "realism",
+                                    default=0.0,
+                                    min=0.0,
+                                    max=1.0,
+                                    step=0.01,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Pulls output toward photographic realism."
+                                    "Leave at 0 for the model default.",
+                                    advanced=True,
+                                ),
+                            ],
+                        ),
+                        IO.DynamicCombo.Option(
+                            "Starlight (Astra) Fast",
+                            [IO.Combo.Input("upscaler_resolution", options=["FullHD (1080p)", "4K (2160p)"]),],
+                        ),
+                        IO.DynamicCombo.Option(
+                            "Starlight (Astra) Creative",
+                            [
+                                IO.Combo.Input("upscaler_resolution", options=["FullHD (1080p)", "4K (2160p)"]),
+                                IO.Combo.Input(
+                                    "creativity",
+                                    options=["low", "middle", "high"],
+                                    default="low",
+                                    tooltip="Creative strength of the upscale.",
+                                ),
+                            ],
+                        ),
+                        IO.DynamicCombo.Option(
+                            "Starlight Precise 2.5",
+                            [IO.Combo.Input("upscaler_resolution", options=["FullHD (1080p)", "4K (2160p)"])],
+                        ),
+                        IO.DynamicCombo.Option("Disabled", []),
+                    ],
+                ),
+                IO.DynamicCombo.Input(
+                    "interpolation_model",
+                    options=[
+                        IO.DynamicCombo.Option("Disabled", []),
+                        IO.DynamicCombo.Option(
+                            "apo-8",
+                            [
+                                IO.Int.Input(
+                                    "interpolation_frame_rate",
+                                    default=60,
+                                    min=15,
+                                    max=240,
+                                    display_mode=IO.NumberDisplay.number,
+                                    tooltip="Output frame rate.",
+                                ),
+                                IO.Int.Input(
+                                    "interpolation_slowmo",
+                                    default=1,
+                                    min=1,
+                                    max=16,
+                                    display_mode=IO.NumberDisplay.number,
+                                    tooltip="Slow-motion factor applied to the input video. "
+                                    "For example, 2 makes the output twice as slow and doubles the duration.",
+                                    advanced=True,
+                                ),
+                                IO.Boolean.Input(
+                                    "interpolation_duplicate",
+                                    default=False,
+                                    tooltip="Analyze the input for duplicate frames and remove them.",
+                                    advanced=True,
+                                ),
+                                IO.Float.Input(
+                                    "interpolation_duplicate_threshold",
+                                    default=0.01,
+                                    min=0.001,
+                                    max=0.1,
+                                    step=0.001,
+                                    display_mode=IO.NumberDisplay.number,
+                                    tooltip="Detection sensitivity for duplicate frames.",
+                                    advanced=True,
+                                ),
+                            ],
+                        ),
+                    ],
+                ),
+                IO.Combo.Input(
+                    "dynamic_compression_level",
+                    options=["Low", "Mid", "High"],
+                    default="Low",
+                    tooltip="CQP level.",
+                    optional=True,
+                ),
+            ],
+            outputs=[
+                IO.Video.Output(),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                depends_on=IO.PriceBadgeDepends(widgets=[
+                    "upscaler_model",
+                    "upscaler_model.upscaler_resolution",
+                    "interpolation_model",
+                ]),
+                expr="""
+                (
+                  $model := $lookup(widgets, "upscaler_model");
+                  $res := $lookup(widgets, "upscaler_model.upscaler_resolution");
+                  $interp := $lookup(widgets, "interpolation_model");
+                  $is4k := $contains($res, "4k");
+                  $hasInterp := $interp != "disabled";
+                  $rates := {
+                    "starlight (astra) fast":     {"hd": 0.43, "uhd": 0.85},
+                    "starlight precise 2.5":      {"hd": 0.70, "uhd": 1.54},
+                    "astra 2":                    {"hd": 1.72, "uhd": 2.85},
+                    "starlight (astra) creative": {"hd": 2.25, "uhd": 3.99}
+                  };
+                  $surcharge := $is4k ? 0.28 : 0.14;
+                  $entry := $lookup($rates, $model);
+                  $base := $is4k ? $entry.uhd : $entry.hd;
+                  $hi := $base + ($hasInterp ? $surcharge : 0);
+                  $model = "disabled"
+                    ? {"type":"text","text":"Interpolation only"}
+                    : ($hasInterp
+                        ? {"type":"text","text":"~" & $string($base) & "–" & $string($hi) & " credits/src frame"}
+                        : {"type":"text","text":"~" & $string($base) & " credits/src frame"})
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        video: Input.Video,
+        upscaler_model: dict,
+        interpolation_model: dict,
+        dynamic_compression_level: str = "Low",
+    ) -> IO.NodeOutput:
+        upscaler_choice = upscaler_model["upscaler_model"]
+        interpolation_choice = interpolation_model["interpolation_model"]
+        if upscaler_choice == "Disabled" and interpolation_choice == "Disabled":
+            raise ValueError("There is nothing to do: both upscaling and interpolation are disabled.")
+        validate_container_format_is_mp4(video)
+        src_width, src_height = video.get_dimensions()
+        src_frame_rate = int(video.get_frame_rate())
+        duration_sec = video.get_duration()
+        src_video_stream = video.get_stream_source()
+        target_width = src_width
+        target_height = src_height
+        target_frame_rate = src_frame_rate
+        filters = []
+        if upscaler_choice != "Disabled":
+            if "1080p" in upscaler_model["upscaler_resolution"]:
+                target_pixel_p = 1080
+                max_long_side = 1920
+            else:
+                target_pixel_p = 2160
+                max_long_side = 3840
+            ar = src_width / src_height
+            if src_width >= src_height:
+                # Landscape or Square; Attempt to set height to target (e.g., 2160), calculate width
+                target_height = target_pixel_p
+                target_width = int(target_height * ar)
+                # Check if width exceeds standard bounds (for ultra-wide e.g., 21:9 ARs)
+                if target_width > max_long_side:
+                    target_width = max_long_side
+                    target_height = int(target_width / ar)
+            else:
+                # Portrait; Attempt to set width to target (e.g., 2160), calculate height
+                target_width = target_pixel_p
+                target_height = int(target_width / ar)
+                # Check if height exceeds standard bounds
+                if target_height > max_long_side:
+                    target_height = max_long_side
+                    target_width = int(target_height * ar)
+            if target_width % 2 != 0:
+                target_width += 1
+            if target_height % 2 != 0:
+                target_height += 1
+            model_id = UPSCALER_MODELS_MAP[upscaler_choice]
+            if model_id == "slc-1":
+                filters.append(
+                    VideoEnhancementFilter(
+                        model=model_id,
+                        creativity=upscaler_model["creativity"],
+                        isOptimizedMode=True,
+                    )
+                )
+            elif model_id == "ast-2":
+                n_frames = video.get_frame_count()
+                ast2_prompt = (upscaler_model["prompt"] or "").strip()
+                if ast2_prompt and n_frames > AST2_MAX_FRAMES_WITH_PROMPT:
+                    raise ValueError(
+                        f"Astra 2 with a prompt is limited to {AST2_MAX_FRAMES_WITH_PROMPT} input frames "
+                        f"(~15s @ 30fps); video has {n_frames}. Clear the prompt or shorten the clip."
+                    )
+                if n_frames > AST2_MAX_FRAMES:
+                    raise ValueError(f"Astra 2 is limited to {AST2_MAX_FRAMES} input frames; video has {n_frames}.")
+                realism = upscaler_model["realism"]
+                filters.append(
+                    VideoEnhancementFilter(
+                        model=model_id,
+                        creativity=upscaler_model["creativity"],
+                        prompt=(ast2_prompt or None),
+                        sharp=upscaler_model["sharp"],
+                        realism=(realism if realism > 0 else None),
+                    )
+                )
+            else:
+                filters.append(VideoEnhancementFilter(model=model_id))
+        if interpolation_choice != "Disabled":
+            target_frame_rate = interpolation_model["interpolation_frame_rate"]
+            filters.append(
+                VideoFrameInterpolationFilter(
+                    model=interpolation_choice,
+                    slowmo=interpolation_model["interpolation_slowmo"],
+                    fps=interpolation_model["interpolation_frame_rate"],
+                    duplicate=interpolation_model["interpolation_duplicate"],
+                    duplicate_threshold=interpolation_model["interpolation_duplicate_threshold"],
+                ),
+            )
+        initial_res = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/topaz/video/", method="POST"),
+            response_model=CreateVideoResponse,
+            data=CreateVideoRequest(
+                source=CreateVideoRequestSource(
+                    container="mp4",
+                    size=get_fs_object_size(src_video_stream),
+                    duration=int(duration_sec),
+                    frameCount=video.get_frame_count(),
+                    frameRate=src_frame_rate,
+                    resolution=Resolution(width=src_width, height=src_height),
+                ),
+                filters=filters,
+                output=OutputInformationVideo(
+                    resolution=Resolution(width=target_width, height=target_height),
+                    frameRate=target_frame_rate,
+                    audioCodec="AAC",
+                    audioTransfer="Copy",
+                    dynamicCompressionLevel=dynamic_compression_level,
+                ),
+            ),
+            wait_label="Creating task",
+            final_label_on_success="Task created",
+        )
+        upload_res = await sync_op(
+            cls,
+            ApiEndpoint(
+                path=f"/proxy/topaz/video/{initial_res.requestId}/accept",
+                method="PATCH",
+            ),
+            response_model=VideoAcceptResponse,
+            wait_label="Preparing upload",
+            final_label_on_success="Upload started",
+        )
+        if len(upload_res.urls) > 1:
+            raise NotImplementedError(
+                "Large files are not currently supported. Please open an issue in the ComfyUI repository."
+            )
+        async with aiohttp.ClientSession(headers={"Content-Type": "video/mp4"}) as session:
+            if isinstance(src_video_stream, BytesIO):
+                src_video_stream.seek(0)
+                async with session.put(upload_res.urls[0], data=src_video_stream, raise_for_status=True) as res:
+                    upload_etag = res.headers["Etag"]
+            else:
+                with builtins.open(src_video_stream, "rb") as video_file:
+                    async with session.put(upload_res.urls[0], data=video_file, raise_for_status=True) as res:
+                        upload_etag = res.headers["Etag"]
+        await sync_op(
+            cls,
+            ApiEndpoint(
+                path=f"/proxy/topaz/video/{initial_res.requestId}/complete-upload",
+                method="PATCH",
+            ),
+            response_model=VideoCompleteUploadResponse,
+            data=VideoCompleteUploadRequest(
+                uploadResults=[
+                    VideoCompleteUploadRequestPart(
+                        partNum=1,
+                        eTag=upload_etag,
+                    ),
+                ],
+            ),
+            wait_label="Finalizing upload",
+            final_label_on_success="Upload completed",
+        )
+        final_response = await poll_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/topaz/video/{initial_res.requestId}/status"),
+            response_model=VideoStatusResponse,
+            status_extractor=lambda x: x.status,
+            progress_extractor=lambda x: getattr(x, "progress", 0),
+            price_extractor=lambda x: (x.estimates.cost[0] * 0.08 if x.estimates and x.estimates.cost[0] else None),
+            poll_interval=10.0,
         )
         return IO.NodeOutput(await download_url_to_video_output(final_response.download.url))
 
@@ -464,6 +819,7 @@ class TopazExtension(ComfyExtension):
         return [
             TopazImageEnhance,
             TopazVideoEnhance,
+            TopazVideoEnhanceV2,
         ]
 
 

comfy_api_nodes/nodes_veo2.py

@@ -393,8 +393,8 @@ class Veo3VideoGenerationNode(IO.ComfyNode):
         model="veo-3.0-generate-001",
         generate_audio=False,
     ):
-        if "lite" in model and resolution == "4k":
-            raise Exception("4K resolution is not supported by the veo-3.1-lite model.")
+        if resolution == "4k" and ("lite" in model or "3.0" in model):
+            raise Exception("4K resolution is not supported by the veo-3.1-lite or veo-3.0 models.")
 
         model = MODELS_MAP[model]
 

comfy_api_nodes/nodes_vidu.py

@@ -38,7 +38,7 @@ async def execute_task(
     cls: type[IO.ComfyNode],
     vidu_endpoint: str,
     payload: TaskCreationRequest | TaskExtendCreationRequest | TaskMultiFrameCreationRequest,
-    max_poll_attempts: int = 320,
+    max_poll_attempts: int = 480,
 ) -> list[TaskResult]:
     task_creation_response = await sync_op(
         cls,
@@ -1097,7 +1097,6 @@ class ViduExtendVideoNode(IO.ComfyNode):
                 video_url=await upload_video_to_comfyapi(cls, video, wait_label="Uploading video"),
                 images=[image_url] if image_url else None,
             ),
-            max_poll_attempts=480,
         )
         return IO.NodeOutput(await download_url_to_video_output(results[0].url))
 

comfy_api_nodes/nodes_wan.py

@@ -818,7 +818,6 @@ class WanReferenceVideoApi(IO.ComfyNode):
             response_model=VideoTaskStatusResponse,
             status_extractor=lambda x: x.output.task_status,
             poll_interval=6,
-            max_poll_attempts=280,
         )
         return IO.NodeOutput(await download_url_to_video_output(response.output.video_url))
 
@@ -1646,6 +1645,557 @@ class Wan2ReferenceVideoApi(IO.ComfyNode):
         return IO.NodeOutput(await download_url_to_video_output(response.output.video_url))
 
 
+class HappyHorseTextToVideoApi(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="HappyHorseTextToVideoApi",
+            display_name="HappyHorse Text to Video",
+            category="api node/video/Wan",
+            description="Generates a video based on a text prompt using the HappyHorse model.",
+            inputs=[
+                IO.DynamicCombo.Input(
+                    "model",
+                    options=[
+                        IO.DynamicCombo.Option(
+                            "happyhorse-1.0-t2v",
+                            [
+                                IO.String.Input(
+                                    "prompt",
+                                    multiline=True,
+                                    default="",
+                                    tooltip="Prompt describing the elements and visual features. "
+                                    "Supports English and Chinese.",
+                                ),
+                                IO.Combo.Input(
+                                    "resolution",
+                                    options=["720P", "1080P"],
+                                ),
+                                IO.Combo.Input(
+                                    "ratio",
+                                    options=["16:9", "9:16", "1:1", "4:3", "3:4"],
+                                ),
+                                IO.Int.Input(
+                                    "duration",
+                                    default=5,
+                                    min=3,
+                                    max=15,
+                                    step=1,
+                                    display_mode=IO.NumberDisplay.number,
+                                ),
+                            ],
+                        ),
+                    ],
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    display_mode=IO.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed to use for generation.",
+                ),
+                IO.Boolean.Input(
+                    "watermark",
+                    default=False,
+                    tooltip="Whether to add an AI-generated watermark to the result.",
+                    advanced=True,
+                ),
+            ],
+            outputs=[
+                IO.Video.Output(),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                depends_on=IO.PriceBadgeDepends(widgets=["model", "model.resolution", "model.duration"]),
+                expr="""
+                (
+                  $res := $lookup(widgets, "model.resolution");
+                  $dur := $lookup(widgets, "model.duration");
+                  $ppsTable := { "720p": 0.14, "1080p": 0.24 };
+                  $pps := $lookup($ppsTable, $res);
+                  { "type": "usd", "usd": $pps * $dur }
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: dict,
+        seed: int,
+        watermark: bool,
+    ):
+        validate_string(model["prompt"], strip_whitespace=False, min_length=1)
+        initial_response = await sync_op(
+            cls,
+            ApiEndpoint(
+                path="/proxy/wan/api/v1/services/aigc/video-generation/video-synthesis",
+                method="POST",
+            ),
+            response_model=TaskCreationResponse,
+            data=Wan27Text2VideoTaskCreationRequest(
+                model=model["model"],
+                input=Text2VideoInputField(
+                    prompt=model["prompt"],
+                    negative_prompt=None,
+                ),
+                parameters=Wan27Text2VideoParametersField(
+                    resolution=model["resolution"],
+                    ratio=model["ratio"],
+                    duration=model["duration"],
+                    seed=seed,
+                    watermark=watermark,
+                ),
+            ),
+        )
+        if not initial_response.output:
+            raise Exception(f"An unknown error occurred: {initial_response.code} - {initial_response.message}")
+        response = await poll_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/wan/api/v1/tasks/{initial_response.output.task_id}"),
+            response_model=VideoTaskStatusResponse,
+            status_extractor=lambda x: x.output.task_status,
+            poll_interval=7,
+        )
+        return IO.NodeOutput(await download_url_to_video_output(response.output.video_url))
+
+
+class HappyHorseImageToVideoApi(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="HappyHorseImageToVideoApi",
+            display_name="HappyHorse Image to Video",
+            category="api node/video/Wan",
+            description="Generate a video from a first-frame image using the HappyHorse model.",
+            inputs=[
+                IO.DynamicCombo.Input(
+                    "model",
+                    options=[
+                        IO.DynamicCombo.Option(
+                            "happyhorse-1.0-i2v",
+                            [
+                                IO.String.Input(
+                                    "prompt",
+                                    multiline=True,
+                                    default="",
+                                    tooltip="Prompt describing the elements and visual features. "
+                                    "Supports English and Chinese.",
+                                ),
+                                IO.Combo.Input(
+                                    "resolution",
+                                    options=["720P", "1080P"],
+                                ),
+                                IO.Int.Input(
+                                    "duration",
+                                    default=5,
+                                    min=3,
+                                    max=15,
+                                    step=1,
+                                    display_mode=IO.NumberDisplay.number,
+                                ),
+                            ],
+                        ),
+                    ],
+                ),
+                IO.Image.Input(
+                    "first_frame",
+                    tooltip="First frame image. The output aspect ratio is derived from this image.",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    display_mode=IO.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed to use for generation.",
+                ),
+                IO.Boolean.Input(
+                    "watermark",
+                    default=False,
+                    tooltip="Whether to add an AI-generated watermark to the result.",
+                    advanced=True,
+                ),
+            ],
+            outputs=[
+                IO.Video.Output(),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                depends_on=IO.PriceBadgeDepends(widgets=["model", "model.resolution", "model.duration"]),
+                expr="""
+                (
+                  $res := $lookup(widgets, "model.resolution");
+                  $dur := $lookup(widgets, "model.duration");
+                  $ppsTable := { "720p": 0.14, "1080p": 0.24 };
+                  $pps := $lookup($ppsTable, $res);
+                  { "type": "usd", "usd": $pps * $dur }
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: dict,
+        first_frame: Input.Image,
+        seed: int,
+        watermark: bool,
+    ):
+        media = [
+            Wan27MediaItem(
+                type="first_frame",
+                url=await upload_image_to_comfyapi(cls, image=first_frame),
+            )
+        ]
+        initial_response = await sync_op(
+            cls,
+            ApiEndpoint(
+                path="/proxy/wan/api/v1/services/aigc/video-generation/video-synthesis",
+                method="POST",
+            ),
+            response_model=TaskCreationResponse,
+            data=Wan27ImageToVideoTaskCreationRequest(
+                model=model["model"],
+                input=Wan27ImageToVideoInputField(
+                    prompt=model["prompt"] or None,
+                    negative_prompt=None,
+                    media=media,
+                ),
+                parameters=Wan27ImageToVideoParametersField(
+                    resolution=model["resolution"],
+                    duration=model["duration"],
+                    seed=seed,
+                    watermark=watermark,
+                ),
+            ),
+        )
+        if not initial_response.output:
+            raise Exception(f"An unknown error occurred: {initial_response.code} - {initial_response.message}")
+        response = await poll_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/wan/api/v1/tasks/{initial_response.output.task_id}"),
+            response_model=VideoTaskStatusResponse,
+            status_extractor=lambda x: x.output.task_status,
+            poll_interval=7,
+        )
+        return IO.NodeOutput(await download_url_to_video_output(response.output.video_url))
+
+
+class HappyHorseVideoEditApi(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="HappyHorseVideoEditApi",
+            display_name="HappyHorse Video Edit",
+            category="api node/video/Wan",
+            description="Edit a video using text instructions or reference images with the HappyHorse model. "
+            "Output duration is 3-15s and matches the input video; inputs longer than 15s are truncated.",
+            inputs=[
+                IO.DynamicCombo.Input(
+                    "model",
+                    options=[
+                        IO.DynamicCombo.Option(
+                            "happyhorse-1.0-video-edit",
+                            [
+                                IO.String.Input(
+                                    "prompt",
+                                    multiline=True,
+                                    default="",
+                                    tooltip="Editing instructions or style transfer requirements.",
+                                ),
+                                IO.Combo.Input(
+                                    "resolution",
+                                    options=["720P", "1080P"],
+                                ),
+                                IO.Combo.Input(
+                                    "ratio",
+                                    options=["16:9", "9:16", "1:1", "4:3", "3:4"],
+                                    tooltip="Aspect ratio. If not changed, approximates the input video ratio.",
+                                ),
+                                IO.Autogrow.Input(
+                                    "reference_images",
+                                    template=IO.Autogrow.TemplateNames(
+                                        IO.Image.Input("reference_image"),
+                                        names=[
+                                            "image1",
+                                            "image2",
+                                            "image3",
+                                            "image4",
+                                            "image5",
+                                        ],
+                                        min=0,
+                                    ),
+                                ),
+                            ],
+                        ),
+                    ],
+                ),
+                IO.Video.Input(
+                    "video",
+                    tooltip="The video to edit.",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    display_mode=IO.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed to use for generation.",
+                ),
+                IO.Boolean.Input(
+                    "watermark",
+                    default=False,
+                    tooltip="Whether to add an AI-generated watermark to the result.",
+                    advanced=True,
+                ),
+            ],
+            outputs=[
+                IO.Video.Output(),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                depends_on=IO.PriceBadgeDepends(widgets=["model", "model.resolution"]),
+                expr="""
+                (
+                  $res := $lookup(widgets, "model.resolution");
+                  $ppsTable := { "720p": 0.14, "1080p": 0.24 };
+                  $pps := $lookup($ppsTable, $res);
+                  { "type": "usd", "usd": $pps, "format": { "suffix": "/second" } }
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: dict,
+        video: Input.Video,
+        seed: int,
+        watermark: bool,
+    ):
+        validate_string(model["prompt"], strip_whitespace=False, min_length=1)
+        validate_video_duration(video, min_duration=3, max_duration=60)
+        media = [Wan27MediaItem(type="video", url=await upload_video_to_comfyapi(cls, video))]
+        reference_images = model.get("reference_images", {})
+        for key in reference_images:
+            media.append(
+                Wan27MediaItem(
+                    type="reference_image", url=await upload_image_to_comfyapi(cls, image=reference_images[key])
+                )
+            )
+        initial_response = await sync_op(
+            cls,
+            ApiEndpoint(
+                path="/proxy/wan/api/v1/services/aigc/video-generation/video-synthesis",
+                method="POST",
+            ),
+            response_model=TaskCreationResponse,
+            data=Wan27VideoEditTaskCreationRequest(
+                model=model["model"],
+                input=Wan27VideoEditInputField(prompt=model["prompt"], media=media),
+                parameters=Wan27VideoEditParametersField(
+                    resolution=model["resolution"],
+                    ratio=model["ratio"],
+                    duration=None,
+                    watermark=watermark,
+                    seed=seed,
+                ),
+            ),
+        )
+        if not initial_response.output:
+            raise Exception(f"An unknown error occurred: {initial_response.code} - {initial_response.message}")
+        response = await poll_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/wan/api/v1/tasks/{initial_response.output.task_id}"),
+            response_model=VideoTaskStatusResponse,
+            status_extractor=lambda x: x.output.task_status,
+            poll_interval=7,
+        )
+        return IO.NodeOutput(await download_url_to_video_output(response.output.video_url))
+
+
+class HappyHorseReferenceVideoApi(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="HappyHorseReferenceVideoApi",
+            display_name="HappyHorse Reference to Video",
+            category="api node/video/Wan",
+            description="Generate a video featuring a person or object from reference materials with the HappyHorse "
+            "model. Supports single-character performances and multi-character interactions.",
+            inputs=[
+                IO.DynamicCombo.Input(
+                    "model",
+                    options=[
+                        IO.DynamicCombo.Option(
+                            "happyhorse-1.0-r2v",
+                            [
+                                IO.String.Input(
+                                    "prompt",
+                                    multiline=True,
+                                    default="",
+                                    tooltip="Prompt describing the video. Use identifiers such as 'character1' and "
+                                    "'character2' to refer to the reference characters.",
+                                ),
+                                IO.Combo.Input(
+                                    "resolution",
+                                    options=["720P", "1080P"],
+                                ),
+                                IO.Combo.Input(
+                                    "ratio",
+                                    options=["16:9", "9:16", "1:1", "4:3", "3:4"],
+                                ),
+                                IO.Int.Input(
+                                    "duration",
+                                    default=5,
+                                    min=3,
+                                    max=15,
+                                    step=1,
+                                    display_mode=IO.NumberDisplay.number,
+                                ),
+                                IO.Autogrow.Input(
+                                    "reference_images",
+                                    template=IO.Autogrow.TemplateNames(
+                                        IO.Image.Input("reference_image"),
+                                        names=[
+                                            "image1",
+                                            "image2",
+                                            "image3",
+                                            "image4",
+                                            "image5",
+                                            "image6",
+                                            "image7",
+                                            "image8",
+                                            "image9",
+                                        ],
+                                        min=1,
+                                    ),
+                                ),
+                            ],
+                        ),
+                    ],
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    display_mode=IO.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed to use for generation.",
+                ),
+                IO.Boolean.Input(
+                    "watermark",
+                    default=False,
+                    tooltip="Whether to add an AI-generated watermark to the result.",
+                    advanced=True,
+                ),
+            ],
+            outputs=[
+                IO.Video.Output(),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                depends_on=IO.PriceBadgeDepends(widgets=["model", "model.resolution", "model.duration"]),
+                expr="""
+                (
+                  $res := $lookup(widgets, "model.resolution");
+                  $dur := $lookup(widgets, "model.duration");
+                  $ppsTable := { "720p": 0.14, "1080p": 0.24 };
+                  $pps := $lookup($ppsTable, $res);
+                  { "type": "usd", "usd": $pps * $dur }
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: dict,
+        seed: int,
+        watermark: bool,
+    ):
+        validate_string(model["prompt"], strip_whitespace=False, min_length=1)
+        media = []
+        reference_images = model.get("reference_images", {})
+        for key in reference_images:
+            media.append(
+                Wan27MediaItem(
+                    type="reference_image",
+                    url=await upload_image_to_comfyapi(cls, image=reference_images[key]),
+                )
+            )
+        if not media:
+            raise ValueError("At least one reference reference image must be provided.")
+
+        initial_response = await sync_op(
+            cls,
+            ApiEndpoint(
+                path="/proxy/wan/api/v1/services/aigc/video-generation/video-synthesis",
+                method="POST",
+            ),
+            response_model=TaskCreationResponse,
+            data=Wan27ReferenceVideoTaskCreationRequest(
+                model=model["model"],
+                input=Wan27ReferenceVideoInputField(
+                    prompt=model["prompt"],
+                    negative_prompt=None,
+                    media=media,
+                ),
+                parameters=Wan27ReferenceVideoParametersField(
+                    resolution=model["resolution"],
+                    ratio=model["ratio"],
+                    duration=model["duration"],
+                    watermark=watermark,
+                    seed=seed,
+                ),
+            ),
+        )
+        if not initial_response.output:
+            raise Exception(f"An unknown error occurred: {initial_response.code} - {initial_response.message}")
+        response = await poll_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/wan/api/v1/tasks/{initial_response.output.task_id}"),
+            response_model=VideoTaskStatusResponse,
+            status_extractor=lambda x: x.output.task_status,
+            poll_interval=7,
+        )
+        return IO.NodeOutput(await download_url_to_video_output(response.output.video_url))
+
+
 class WanApiExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[IO.ComfyNode]]:
@@ -1660,6 +2210,10 @@ class WanApiExtension(ComfyExtension):
             Wan2VideoContinuationApi,
             Wan2VideoEditApi,
             Wan2ReferenceVideoApi,
+            HappyHorseTextToVideoApi,
+            HappyHorseImageToVideoApi,
+            HappyHorseVideoEditApi,
+            HappyHorseReferenceVideoApi,
         ]
 
 

comfy_api_nodes/nodes_wavespeed.py

@@ -84,7 +84,6 @@ class WavespeedFlashVSRNode(IO.ComfyNode):
             response_model=TaskResultResponse,
             status_extractor=lambda x: "failed" if x.data is None else x.data.status,
             poll_interval=10.0,
-            max_poll_attempts=480,
         )
         if final_response.code != 200:
             raise ValueError(
@@ -156,7 +155,6 @@ class WavespeedImageUpscaleNode(IO.ComfyNode):
             response_model=TaskResultResponse,
             status_extractor=lambda x: "failed" if x.data is None else x.data.status,
             poll_interval=10.0,
-            max_poll_attempts=480,
         )
         if final_response.code != 200:
             raise ValueError(

comfy_api_nodes/util/client.py

@@ -150,7 +150,7 @@ async def poll_op(
     queued_statuses: list[str | int] | None = None,
     data: BaseModel | None = None,
     poll_interval: float = 5.0,
-    max_poll_attempts: int = 160,
+    max_poll_attempts: int = 480,
     timeout_per_poll: float = 120.0,
     max_retries_per_poll: int = 10,
     retry_delay_per_poll: float = 1.0,
@@ -158,6 +158,7 @@ async def poll_op(
     estimated_duration: int | None = None,
     cancel_endpoint: ApiEndpoint | None = None,
     cancel_timeout: float = 10.0,
+    extra_text: str | None = None,
 ) -> M:
     raw = await poll_op_raw(
         cls,
@@ -178,6 +179,7 @@ async def poll_op(
         estimated_duration=estimated_duration,
         cancel_endpoint=cancel_endpoint,
         cancel_timeout=cancel_timeout,
+        extra_text=extra_text,
     )
     if not isinstance(raw, dict):
         raise Exception("Expected JSON response to validate into a Pydantic model, got non-JSON (binary or text).")
@@ -254,7 +256,7 @@ async def poll_op_raw(
     queued_statuses: list[str | int] | None = None,
     data: dict[str, Any] | BaseModel | None = None,
     poll_interval: float = 5.0,
-    max_poll_attempts: int = 160,
+    max_poll_attempts: int = 480,
     timeout_per_poll: float = 120.0,
     max_retries_per_poll: int = 10,
     retry_delay_per_poll: float = 1.0,
@@ -262,6 +264,7 @@ async def poll_op_raw(
     estimated_duration: int | None = None,
     cancel_endpoint: ApiEndpoint | None = None,
     cancel_timeout: float = 10.0,
+    extra_text: str | None = None,
 ) -> dict[str, Any]:
     """
     Polls an endpoint until the task reaches a terminal state. Displays time while queued/processing,
@@ -301,6 +304,7 @@ async def poll_op_raw(
                     price=state.price,
                     is_queued=state.is_queued,
                     processing_elapsed_seconds=int(proc_elapsed),
+                    extra_text=extra_text,
                 )
                 await asyncio.sleep(1.0)
         except Exception as exc:
@@ -391,6 +395,7 @@ async def poll_op_raw(
                     price=state.price,
                     is_queued=False,
                     processing_elapsed_seconds=int(state.base_processing_elapsed),
+                    extra_text=extra_text,
                 )
                 return resp_json
 
@@ -464,6 +469,7 @@ def _display_time_progress(
     price: float | None = None,
     is_queued: bool | None = None,
     processing_elapsed_seconds: int | None = None,
+    extra_text: str | None = None,
 ) -> None:
     if estimated_total is not None and estimated_total > 0 and is_queued is False:
         pe = processing_elapsed_seconds if processing_elapsed_seconds is not None else elapsed_seconds
@@ -471,7 +477,8 @@ def _display_time_progress(
         time_line = f"Time elapsed: {int(elapsed_seconds)}s (~{remaining}s remaining)"
     else:
         time_line = f"Time elapsed: {int(elapsed_seconds)}s"
-    _display_text(node_cls, time_line, status=status, price=price)
+    text = f"{time_line}\n\n{extra_text}" if extra_text else time_line
+    _display_text(node_cls, text, status=status, price=price)
 
 
 async def _diagnose_connectivity() -> dict[str, bool]:

comfy_execution/caching.py

@@ -5,6 +5,7 @@ import psutil
 import time
 import torch
 from typing import Sequence, Mapping, Dict
+from comfy.model_patcher import ModelPatcher
 from comfy_execution.graph import DynamicPrompt
 from abc import ABC, abstractmethod
 
@@ -523,13 +524,15 @@ class RAMPressureCache(LRUCache):
         self.timestamps[self.cache_key_set.get_data_key(node_id)] = time.time()
         super().set_local(node_id, value)
 
-    def ram_release(self, target):
+    def ram_release(self, target, free_active=False):
         if psutil.virtual_memory().available >= target:
             return
 
         clean_list = []
 
         for key, cache_entry in self.cache.items():
+            if not free_active and self.used_generation[key] == self.generation:
+                continue
             oom_score =  RAM_CACHE_OLD_WORKFLOW_OOM_MULTIPLIER ** (self.generation - self.used_generation[key])
 
             ram_usage = RAM_CACHE_DEFAULT_RAM_USAGE
@@ -542,6 +545,9 @@ class RAMPressureCache(LRUCache):
                         scan_list_for_ram_usage(output)
                     elif isinstance(output, torch.Tensor) and output.device.type == 'cpu':
                         ram_usage += output.numel() * output.element_size()
+                    elif isinstance(output, ModelPatcher) and self.used_generation[key] != self.generation:
+                        #old ModelPatchers are the first to go
+                        ram_usage = 1e30
             scan_list_for_ram_usage(cache_entry.outputs)
 
             oom_score *= ram_usage

comfy_extras/frame_interpolation_models/film_net.py

@@ -0,0 +1,258 @@
+"""FILM: Frame Interpolation for Large Motion (ECCV 2022)."""
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+import comfy.ops
+
+ops = comfy.ops.disable_weight_init
+
+
+class FilmConv2d(nn.Module):
+    """Conv2d with optional LeakyReLU and FILM-style padding."""
+
+    def __init__(self, in_channels, out_channels, size, activation=True, device=None, dtype=None, operations=ops):
+        super().__init__()
+        self.even_pad = not size % 2
+        self.conv = operations.Conv2d(in_channels, out_channels, kernel_size=size, padding=size // 2 if size % 2 else 0, device=device, dtype=dtype)
+        self.activation = nn.LeakyReLU(0.2) if activation else None
+
+    def forward(self, x):
+        if self.even_pad:
+            x = F.pad(x, (0, 1, 0, 1))
+        x = self.conv(x)
+        if self.activation is not None:
+            x = self.activation(x)
+        return x
+
+
+def _warp_core(image, flow, grid_x, grid_y):
+    dtype = image.dtype
+    H, W = flow.shape[2], flow.shape[3]
+    dx = flow[:, 0].float() / (W * 0.5)
+    dy = flow[:, 1].float() / (H * 0.5)
+    grid = torch.stack([grid_x[None, None, :] + dx, grid_y[None, :, None] + dy], dim=3)
+    return F.grid_sample(image.float(), grid, mode="bilinear", padding_mode="border", align_corners=False).to(dtype)
+
+
+def build_image_pyramid(image, pyramid_levels):
+    pyramid = [image]
+    for _ in range(1, pyramid_levels):
+        image = F.avg_pool2d(image, 2, 2)
+        pyramid.append(image)
+    return pyramid
+
+
+def flow_pyramid_synthesis(residual_pyramid):
+    flow = residual_pyramid[-1]
+    flow_pyramid = [flow]
+    for residual_flow in residual_pyramid[:-1][::-1]:
+        flow = F.interpolate(flow, size=residual_flow.shape[2:4], mode="bilinear", scale_factor=None).mul_(2).add_(residual_flow)
+        flow_pyramid.append(flow)
+    flow_pyramid.reverse()
+    return flow_pyramid
+
+
+def multiply_pyramid(pyramid, scalar):
+    return [image * scalar[:, None, None, None] for image in pyramid]
+
+
+def pyramid_warp(feature_pyramid, flow_pyramid, warp_fn):
+    return [warp_fn(features, flow) for features, flow in zip(feature_pyramid, flow_pyramid)]
+
+
+def concatenate_pyramids(pyramid1, pyramid2):
+    return [torch.cat([f1, f2], dim=1) for f1, f2 in zip(pyramid1, pyramid2)]
+
+
+class SubTreeExtractor(nn.Module):
+    def __init__(self, in_channels=3, channels=64, n_layers=4, device=None, dtype=None, operations=ops):
+        super().__init__()
+        convs = []
+        for i in range(n_layers):
+            out_ch = channels << i
+            convs.append(nn.Sequential(
+                FilmConv2d(in_channels, out_ch, 3, device=device, dtype=dtype, operations=operations),
+                FilmConv2d(out_ch, out_ch, 3, device=device, dtype=dtype, operations=operations)))
+            in_channels = out_ch
+        self.convs = nn.ModuleList(convs)
+
+    def forward(self, image, n):
+        head = image
+        pyramid = []
+        for i, layer in enumerate(self.convs):
+            head = layer(head)
+            pyramid.append(head)
+            if i < n - 1:
+                head = F.avg_pool2d(head, 2, 2)
+        return pyramid
+
+
+class FeatureExtractor(nn.Module):
+    def __init__(self, in_channels=3, channels=64, sub_levels=4, device=None, dtype=None, operations=ops):
+        super().__init__()
+        self.extract_sublevels = SubTreeExtractor(in_channels, channels, sub_levels, device=device, dtype=dtype, operations=operations)
+        self.sub_levels = sub_levels
+
+    def forward(self, image_pyramid):
+        sub_pyramids = [self.extract_sublevels(image_pyramid[i], min(len(image_pyramid) - i, self.sub_levels))
+                        for i in range(len(image_pyramid))]
+        feature_pyramid = []
+        for i in range(len(image_pyramid)):
+            features = sub_pyramids[i][0]
+            for j in range(1, self.sub_levels):
+                if j <= i:
+                    features = torch.cat([features, sub_pyramids[i - j][j]], dim=1)
+            feature_pyramid.append(features)
+            # Free sub-pyramids no longer needed by future levels
+            if i >= self.sub_levels - 1:
+                sub_pyramids[i - self.sub_levels + 1] = None
+        return feature_pyramid
+
+
+class FlowEstimator(nn.Module):
+    def __init__(self, in_channels, num_convs, num_filters, device=None, dtype=None, operations=ops):
+        super().__init__()
+        self._convs = nn.ModuleList()
+        for _ in range(num_convs):
+            self._convs.append(FilmConv2d(in_channels, num_filters, 3, device=device, dtype=dtype, operations=operations))
+            in_channels = num_filters
+        self._convs.append(FilmConv2d(in_channels, num_filters // 2, 1, device=device, dtype=dtype, operations=operations))
+        self._convs.append(FilmConv2d(num_filters // 2, 2, 1, activation=False, device=device, dtype=dtype, operations=operations))
+
+    def forward(self, features_a, features_b):
+        net = torch.cat([features_a, features_b], dim=1)
+        for conv in self._convs:
+            net = conv(net)
+        return net
+
+
+class PyramidFlowEstimator(nn.Module):
+    def __init__(self, filters=64, flow_convs=(3, 3, 3, 3), flow_filters=(32, 64, 128, 256), device=None, dtype=None, operations=ops):
+        super().__init__()
+        in_channels = filters << 1
+        predictors = []
+        for i in range(len(flow_convs)):
+            predictors.append(FlowEstimator(in_channels, flow_convs[i], flow_filters[i], device=device, dtype=dtype, operations=operations))
+            in_channels += filters << (i + 2)
+        self._predictor = predictors[-1]
+        self._predictors = nn.ModuleList(predictors[:-1][::-1])
+
+    def forward(self, feature_pyramid_a, feature_pyramid_b, warp_fn):
+        levels = len(feature_pyramid_a)
+        v = self._predictor(feature_pyramid_a[-1], feature_pyramid_b[-1])
+        residuals = [v]
+        # Coarse-to-fine: shared predictor for deep levels, then specialized predictors for fine levels
+        steps = [(i, self._predictor) for i in range(levels - 2, len(self._predictors) - 1, -1)]
+        steps += [(len(self._predictors) - 1 - k, p) for k, p in enumerate(self._predictors)]
+        for i, predictor in steps:
+            v = F.interpolate(v, size=feature_pyramid_a[i].shape[2:4], mode="bilinear").mul_(2)
+            v_residual = predictor(feature_pyramid_a[i], warp_fn(feature_pyramid_b[i], v))
+            residuals.append(v_residual)
+            v = v.add_(v_residual)
+        residuals.reverse()
+        return residuals
+
+
+def _get_fusion_channels(level, filters):
+    # Per direction: multi-scale features + RGB image (3ch) + flow (2ch), doubled for both directions
+    return (sum(filters << i for i in range(level)) + 3 + 2) * 2
+
+
+class Fusion(nn.Module):
+    def __init__(self, n_layers=4, specialized_layers=3, filters=64, device=None, dtype=None, operations=ops):
+        super().__init__()
+        self.output_conv = operations.Conv2d(filters, 3, kernel_size=1, device=device, dtype=dtype)
+        self.convs = nn.ModuleList()
+        in_channels = _get_fusion_channels(n_layers, filters)
+        increase = 0
+        for i in range(n_layers)[::-1]:
+            num_filters = (filters << i) if i < specialized_layers else (filters << specialized_layers)
+            self.convs.append(nn.ModuleList([
+                FilmConv2d(in_channels, num_filters, 2, activation=False, device=device, dtype=dtype, operations=operations),
+                FilmConv2d(in_channels + (increase or num_filters), num_filters, 3, device=device, dtype=dtype, operations=operations),
+                FilmConv2d(num_filters, num_filters, 3, device=device, dtype=dtype, operations=operations)]))
+            in_channels = num_filters
+            increase = _get_fusion_channels(i, filters) - num_filters // 2
+
+    def forward(self, pyramid):
+        net = pyramid[-1]
+        for k, layers in enumerate(self.convs):
+            i = len(self.convs) - 1 - k
+            net = layers[0](F.interpolate(net, size=pyramid[i].shape[2:4], mode="nearest"))
+            net = layers[2](layers[1](torch.cat([pyramid[i], net], dim=1)))
+        return self.output_conv(net)
+
+
+class FILMNet(nn.Module):
+    def __init__(self, pyramid_levels=7, fusion_pyramid_levels=5, specialized_levels=3, sub_levels=4,
+                 filters=64, flow_convs=(3, 3, 3, 3), flow_filters=(32, 64, 128, 256), device=None, dtype=None, operations=ops):
+        super().__init__()
+        self.pyramid_levels = pyramid_levels
+        self.fusion_pyramid_levels = fusion_pyramid_levels
+        self.extract = FeatureExtractor(3, filters, sub_levels, device=device, dtype=dtype, operations=operations)
+        self.predict_flow = PyramidFlowEstimator(filters, flow_convs, flow_filters, device=device, dtype=dtype, operations=operations)
+        self.fuse = Fusion(sub_levels, specialized_levels, filters, device=device, dtype=dtype, operations=operations)
+        self._warp_grids = {}
+
+    def get_dtype(self):
+        return self.extract.extract_sublevels.convs[0][0].conv.weight.dtype
+
+    def _build_warp_grids(self, H, W, device):
+        """Pre-compute warp grids for all pyramid levels."""
+        if (H, W) in self._warp_grids:
+            return
+        self._warp_grids = {}  # clear old resolution grids to prevent memory leaks
+        for _ in range(self.pyramid_levels):
+            self._warp_grids[(H, W)] = (
+                torch.linspace(-(1 - 1 / W), 1 - 1 / W, W, dtype=torch.float32, device=device),
+                torch.linspace(-(1 - 1 / H), 1 - 1 / H, H, dtype=torch.float32, device=device),
+            )
+            H, W = H // 2, W // 2
+
+    def warp(self, image, flow):
+        grid_x, grid_y = self._warp_grids[(flow.shape[2], flow.shape[3])]
+        return _warp_core(image, flow, grid_x, grid_y)
+
+    def extract_features(self, img):
+        """Extract image and feature pyramids for a single frame. Can be cached across pairs."""
+        image_pyramid = build_image_pyramid(img, self.pyramid_levels)
+        feature_pyramid = self.extract(image_pyramid)
+        return image_pyramid, feature_pyramid
+
+    def forward(self, img0, img1, timestep=0.5, cache=None):
+        # FILM uses a scalar timestep per batch element (spatially-varying timesteps not supported)
+        t = timestep.mean(dim=(1, 2, 3)).item() if isinstance(timestep, torch.Tensor) else timestep
+        return self.forward_multi_timestep(img0, img1, [t], cache=cache)
+
+    def forward_multi_timestep(self, img0, img1, timesteps, cache=None):
+        """Compute flow once, synthesize at multiple timesteps. Expects batch=1 inputs."""
+        self._build_warp_grids(img0.shape[2], img0.shape[3], img0.device)
+
+        image_pyr0, feat_pyr0 = cache["img0"] if cache and "img0" in cache else self.extract_features(img0)
+        image_pyr1, feat_pyr1 = cache["img1"] if cache and "img1" in cache else self.extract_features(img1)
+
+        fwd_flow = flow_pyramid_synthesis(self.predict_flow(feat_pyr0, feat_pyr1, self.warp))[:self.fusion_pyramid_levels]
+        bwd_flow = flow_pyramid_synthesis(self.predict_flow(feat_pyr1, feat_pyr0, self.warp))[:self.fusion_pyramid_levels]
+
+        # Build warp targets and free full pyramids (only first fpl levels needed from here)
+        fpl = self.fusion_pyramid_levels
+        p2w = [concatenate_pyramids(image_pyr0[:fpl], feat_pyr0[:fpl]),
+               concatenate_pyramids(image_pyr1[:fpl], feat_pyr1[:fpl])]
+        del image_pyr0, image_pyr1, feat_pyr0, feat_pyr1
+
+        results = []
+        dt_tensors = torch.tensor(timesteps, device=img0.device, dtype=img0.dtype)
+        for idx in range(len(timesteps)):
+            batch_dt = dt_tensors[idx:idx + 1]
+            bwd_scaled = multiply_pyramid(bwd_flow, batch_dt)
+            fwd_scaled = multiply_pyramid(fwd_flow, 1 - batch_dt)
+            fwd_warped = pyramid_warp(p2w[0], bwd_scaled, self.warp)
+            bwd_warped = pyramid_warp(p2w[1], fwd_scaled, self.warp)
+            aligned = [torch.cat([fw, bw, bf, ff], dim=1)
+                       for fw, bw, bf, ff in zip(fwd_warped, bwd_warped, bwd_scaled, fwd_scaled)]
+            del fwd_warped, bwd_warped, bwd_scaled, fwd_scaled
+            results.append(self.fuse(aligned))
+            del aligned
+        return torch.cat(results, dim=0)

comfy_extras/frame_interpolation_models/ifnet.py

@@ -0,0 +1,128 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+import comfy.ops
+
+ops = comfy.ops.disable_weight_init
+
+
+def _warp(img, flow, warp_grids):
+    B, _, H, W = img.shape
+    base_grid, flow_div = warp_grids[(H, W)]
+    flow_norm = torch.cat([flow[:, 0:1] / flow_div[0], flow[:, 1:2] / flow_div[1]], 1).float()
+    grid = (base_grid.expand(B, -1, -1, -1) + flow_norm).permute(0, 2, 3, 1)
+    return F.grid_sample(img.float(), grid, mode="bilinear", padding_mode="border", align_corners=True).to(img.dtype)
+
+
+class Head(nn.Module):
+    def __init__(self, out_ch=4, device=None, dtype=None, operations=ops):
+        super().__init__()
+        self.cnn0 = operations.Conv2d(3, 16, 3, 2, 1, device=device, dtype=dtype)
+        self.cnn1 = operations.Conv2d(16, 16, 3, 1, 1, device=device, dtype=dtype)
+        self.cnn2 = operations.Conv2d(16, 16, 3, 1, 1, device=device, dtype=dtype)
+        self.cnn3 = operations.ConvTranspose2d(16, out_ch, 4, 2, 1, device=device, dtype=dtype)
+        self.relu = nn.LeakyReLU(0.2, True)
+
+    def forward(self, x):
+        x = self.relu(self.cnn0(x))
+        x = self.relu(self.cnn1(x))
+        x = self.relu(self.cnn2(x))
+        return self.cnn3(x)
+
+
+class ResConv(nn.Module):
+    def __init__(self, c, device=None, dtype=None, operations=ops):
+        super().__init__()
+        self.conv = operations.Conv2d(c, c, 3, 1, 1, device=device, dtype=dtype)
+        self.beta = nn.Parameter(torch.ones((1, c, 1, 1), device=device, dtype=dtype))
+        self.relu = nn.LeakyReLU(0.2, True)
+
+    def forward(self, x):
+        return self.relu(torch.addcmul(x, self.conv(x), self.beta))
+
+
+class IFBlock(nn.Module):
+    def __init__(self, in_planes, c=64, device=None, dtype=None, operations=ops):
+        super().__init__()
+        self.conv0 = nn.Sequential(
+            nn.Sequential(operations.Conv2d(in_planes, c // 2, 3, 2, 1, device=device, dtype=dtype), nn.LeakyReLU(0.2, True)),
+            nn.Sequential(operations.Conv2d(c // 2, c, 3, 2, 1, device=device, dtype=dtype), nn.LeakyReLU(0.2, True)))
+        self.convblock = nn.Sequential(*(ResConv(c, device=device, dtype=dtype, operations=operations) for _ in range(8)))
+        self.lastconv = nn.Sequential(operations.ConvTranspose2d(c, 4 * 13, 4, 2, 1, device=device, dtype=dtype), nn.PixelShuffle(2))
+
+    def forward(self, x, flow=None, scale=1):
+        x = F.interpolate(x, scale_factor=1.0 / scale, mode="bilinear")
+        if flow is not None:
+            flow = F.interpolate(flow, scale_factor=1.0 / scale, mode="bilinear").div_(scale)
+            x = torch.cat((x, flow), 1)
+        feat = self.convblock(self.conv0(x))
+        tmp = F.interpolate(self.lastconv(feat), scale_factor=scale, mode="bilinear")
+        return tmp[:, :4] * scale, tmp[:, 4:5], tmp[:, 5:]
+
+
+class IFNet(nn.Module):
+    def __init__(self, head_ch=4, channels=(192, 128, 96, 64, 32), device=None, dtype=None, operations=ops):
+        super().__init__()
+        self.encode = Head(out_ch=head_ch, device=device, dtype=dtype, operations=operations)
+        block_in = [7 + 2 * head_ch] + [8 + 4 + 8 + 2 * head_ch] * 4
+        self.blocks = nn.ModuleList([IFBlock(block_in[i], channels[i], device=device, dtype=dtype, operations=operations) for i in range(5)])
+        self.scale_list = [16, 8, 4, 2, 1]
+        self.pad_align = 64
+        self._warp_grids = {}
+
+    def get_dtype(self):
+        return self.encode.cnn0.weight.dtype
+
+    def _build_warp_grids(self, H, W, device):
+        if (H, W) in self._warp_grids:
+            return
+        self._warp_grids = {}  # clear old resolution grids to prevent memory leaks
+        grid_y, grid_x = torch.meshgrid(
+            torch.linspace(-1.0, 1.0, H, device=device, dtype=torch.float32),
+            torch.linspace(-1.0, 1.0, W, device=device, dtype=torch.float32), indexing="ij")
+        self._warp_grids[(H, W)] = (
+            torch.stack((grid_x, grid_y), dim=0).unsqueeze(0),
+            torch.tensor([(W - 1.0) / 2.0, (H - 1.0) / 2.0], dtype=torch.float32, device=device))
+
+    def warp(self, img, flow):
+        return _warp(img, flow, self._warp_grids)
+
+    def extract_features(self, img):
+        """Extract head features for a single frame. Can be cached across pairs."""
+        return self.encode(img)
+
+    def forward(self, img0, img1, timestep=0.5, cache=None):
+        if not isinstance(timestep, torch.Tensor):
+            timestep = torch.full((img0.shape[0], 1, img0.shape[2], img0.shape[3]), timestep, device=img0.device, dtype=img0.dtype)
+
+        self._build_warp_grids(img0.shape[2], img0.shape[3], img0.device)
+
+        B = img0.shape[0]
+        f0 = cache["img0"].expand(B, -1, -1, -1) if cache and "img0" in cache else self.encode(img0)
+        f1 = cache["img1"].expand(B, -1, -1, -1) if cache and "img1" in cache else self.encode(img1)
+        flow = mask = feat = None
+        warped_img0, warped_img1 = img0, img1
+        for i, block in enumerate(self.blocks):
+            if flow is None:
+                flow, mask, feat = block(torch.cat((img0, img1, f0, f1, timestep), 1), None, scale=self.scale_list[i])
+            else:
+                fd, mask, feat = block(
+                    torch.cat((warped_img0, warped_img1, self.warp(f0, flow[:, :2]), self.warp(f1, flow[:, 2:4]), timestep, mask, feat), 1),
+                    flow, scale=self.scale_list[i])
+                flow = flow.add_(fd)
+            warped_img0 = self.warp(img0, flow[:, :2])
+            warped_img1 = self.warp(img1, flow[:, 2:4])
+        return torch.lerp(warped_img1, warped_img0, torch.sigmoid(mask))
+
+
+def detect_rife_config(state_dict):
+    head_ch = state_dict["encode.cnn3.weight"].shape[1]  # ConvTranspose2d: (in_ch, out_ch, kH, kW)
+    channels = []
+    for i in range(5):
+        key = f"blocks.{i}.conv0.1.0.weight"
+        if key in state_dict:
+            channels.append(state_dict[key].shape[0])
+    if len(channels) != 5:
+        raise ValueError(f"Unsupported RIFE model: expected 5 blocks, found {len(channels)}")
+    return head_ch, channels

comfy_extras/nodes_compositing.py

@@ -202,14 +202,11 @@ class JoinImageWithAlpha(io.ComfyNode):
 
     @classmethod
     def execute(cls, image: torch.Tensor, alpha: torch.Tensor) -> io.NodeOutput:
-        batch_size = min(len(image), len(alpha))
-        out_images = []
-
+        batch_size = max(len(image), len(alpha))
         alpha = 1.0 - resize_mask(alpha, image.shape[1:])
-        for i in range(batch_size):
-           out_images.append(torch.cat((image[i][:,:,:3], alpha[i].unsqueeze(2)), dim=2))
-
-        return io.NodeOutput(torch.stack(out_images))
+        alpha = comfy.utils.repeat_to_batch_size(alpha, batch_size)
+        image = comfy.utils.repeat_to_batch_size(image, batch_size)
+        return io.NodeOutput(torch.cat((image[..., :3], alpha.unsqueeze(-1)), dim=-1))
 
 
 class CompositingExtension(ComfyExtension):

comfy_extras/nodes_frame_interpolation.py

@@ -0,0 +1,211 @@
+import torch
+from tqdm import tqdm
+from typing_extensions import override
+
+import comfy.model_patcher
+import comfy.utils
+import folder_paths
+from comfy import model_management
+from comfy_extras.frame_interpolation_models.ifnet import IFNet, detect_rife_config
+from comfy_extras.frame_interpolation_models.film_net import FILMNet
+from comfy_api.latest import ComfyExtension, io
+
+FrameInterpolationModel = io.Custom("INTERP_MODEL")
+
+
+class FrameInterpolationModelLoader(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="FrameInterpolationModelLoader",
+            display_name="Load Frame Interpolation Model",
+            category="loaders",
+            inputs=[
+                io.Combo.Input("model_name", options=folder_paths.get_filename_list("frame_interpolation"),
+                               tooltip="Select a frame interpolation model to load. Models must be placed in the 'frame_interpolation' folder."),
+            ],
+            outputs=[
+                FrameInterpolationModel.Output(),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, model_name) -> io.NodeOutput:
+        model_path = folder_paths.get_full_path_or_raise("frame_interpolation", model_name)
+        sd = comfy.utils.load_torch_file(model_path, safe_load=True)
+
+        model = cls._detect_and_load(sd)
+        dtype = torch.float16 if model_management.should_use_fp16(model_management.get_torch_device()) else torch.float32
+        model.eval().to(dtype)
+        patcher = comfy.model_patcher.ModelPatcher(
+            model,
+            load_device=model_management.get_torch_device(),
+            offload_device=model_management.unet_offload_device(),
+        )
+        return io.NodeOutput(patcher)
+
+    @classmethod
+    def _detect_and_load(cls, sd):
+        # Try FILM
+        if "extract.extract_sublevels.convs.0.0.conv.weight" in sd:
+            model = FILMNet()
+            model.load_state_dict(sd)
+            return model
+
+        # Try RIFE (needs key remapping for raw checkpoints)
+        sd = comfy.utils.state_dict_prefix_replace(sd, {"module.": "", "flownet.": ""})
+        key_map = {}
+        for k in sd:
+            for i in range(5):
+                if k.startswith(f"block{i}."):
+                    key_map[k] = f"blocks.{i}.{k[len(f'block{i}.'):]}"
+        if key_map:
+            sd = {key_map.get(k, k): v for k, v in sd.items()}
+        sd = {k: v for k, v in sd.items() if not k.startswith(("teacher.", "caltime."))}
+
+        try:
+            head_ch, channels = detect_rife_config(sd)
+        except (KeyError, ValueError):
+            raise ValueError("Unrecognized frame interpolation model format")
+        model = IFNet(head_ch=head_ch, channels=channels)
+        model.load_state_dict(sd)
+        return model
+
+
+class FrameInterpolate(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="FrameInterpolate",
+            display_name="Frame Interpolate",
+            category="image/video",
+            search_aliases=["rife", "film", "frame interpolation", "slow motion", "interpolate frames", "vfi"],
+            inputs=[
+                FrameInterpolationModel.Input("interp_model"),
+                io.Image.Input("images"),
+                io.Int.Input("multiplier", default=2, min=2, max=16),
+            ],
+            outputs=[
+                io.Image.Output(),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, interp_model, images, multiplier) -> io.NodeOutput:
+        offload_device = model_management.intermediate_device()
+
+        num_frames = images.shape[0]
+        if num_frames < 2 or multiplier < 2:
+            return io.NodeOutput(images)
+
+        model_management.load_model_gpu(interp_model)
+        device = interp_model.load_device
+        dtype = interp_model.model_dtype()
+        inference_model = interp_model.model
+
+        # Free VRAM for inference activations (model weights + ~20x a single frame's worth)
+        H, W = images.shape[1], images.shape[2]
+        activation_mem = H * W * 3 * images.element_size() * 20
+        model_management.free_memory(activation_mem, device)
+        align = getattr(inference_model, "pad_align", 1)
+
+        # Prepare a single padded frame on device for determining output dimensions
+        def prepare_frame(idx):
+            frame = images[idx:idx + 1].movedim(-1, 1).to(dtype=dtype, device=device)
+            if align > 1:
+                from comfy.ldm.common_dit import pad_to_patch_size
+                frame = pad_to_patch_size(frame, (align, align), padding_mode="reflect")
+            return frame
+
+        # Count total interpolation passes for progress bar
+        total_pairs = num_frames - 1
+        num_interp = multiplier - 1
+        total_steps = total_pairs * num_interp
+        pbar = comfy.utils.ProgressBar(total_steps)
+        tqdm_bar = tqdm(total=total_steps, desc="Frame interpolation")
+
+        batch = num_interp  # reduced on OOM and persists across pairs (same resolution = same limit)
+        t_values = [t / multiplier for t in range(1, multiplier)]
+
+        out_dtype = model_management.intermediate_dtype()
+        total_out_frames = total_pairs * multiplier + 1
+        result = torch.empty((total_out_frames, 3, H, W), dtype=out_dtype, device=offload_device)
+        result[0] = images[0].movedim(-1, 0).to(out_dtype)
+        out_idx = 1
+
+        # Pre-compute timestep tensor on device (padded dimensions needed)
+        sample = prepare_frame(0)
+        pH, pW = sample.shape[2], sample.shape[3]
+        ts_full = torch.tensor(t_values, device=device, dtype=dtype).reshape(num_interp, 1, 1, 1)
+        ts_full = ts_full.expand(-1, 1, pH, pW)
+        del sample
+
+        multi_fn = getattr(inference_model, "forward_multi_timestep", None)
+        feat_cache = {}
+        prev_frame = None
+
+        try:
+            for i in range(total_pairs):
+                img0_single = prev_frame if prev_frame is not None else prepare_frame(i)
+                img1_single = prepare_frame(i + 1)
+                prev_frame = img1_single
+
+                # Cache features: img1 of pair N becomes img0 of pair N+1
+                feat_cache["img0"] = feat_cache.pop("next") if "next" in feat_cache else inference_model.extract_features(img0_single)
+                feat_cache["img1"] = inference_model.extract_features(img1_single)
+                feat_cache["next"] = feat_cache["img1"]
+
+                used_multi = False
+                if multi_fn is not None:
+                    # Models with timestep-independent flow can compute it once for all timesteps
+                    try:
+                        mids = multi_fn(img0_single, img1_single, t_values, cache=feat_cache)
+                        result[out_idx:out_idx + num_interp] = mids[:, :, :H, :W].to(out_dtype)
+                        out_idx += num_interp
+                        pbar.update(num_interp)
+                        tqdm_bar.update(num_interp)
+                        used_multi = True
+                    except model_management.OOM_EXCEPTION:
+                        model_management.soft_empty_cache()
+                        multi_fn = None  # fall through to single-timestep path
+
+                if not used_multi:
+                    j = 0
+                    while j < num_interp:
+                        b = min(batch, num_interp - j)
+                        try:
+                            img0 = img0_single.expand(b, -1, -1, -1)
+                            img1 = img1_single.expand(b, -1, -1, -1)
+                            mids = inference_model(img0, img1, timestep=ts_full[j:j + b], cache=feat_cache)
+                            result[out_idx:out_idx + b] = mids[:, :, :H, :W].to(out_dtype)
+                            out_idx += b
+                            pbar.update(b)
+                            tqdm_bar.update(b)
+                            j += b
+                        except model_management.OOM_EXCEPTION:
+                            if batch <= 1:
+                                raise
+                            batch = max(1, batch // 2)
+                            model_management.soft_empty_cache()
+
+                result[out_idx] = images[i + 1].movedim(-1, 0).to(out_dtype)
+                out_idx += 1
+        finally:
+            tqdm_bar.close()
+
+        # BCHW -> BHWC
+        result = result.movedim(1, -1).clamp_(0.0, 1.0)
+        return io.NodeOutput(result)
+
+
+class FrameInterpolationExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            FrameInterpolationModelLoader,
+            FrameInterpolate,
+        ]
+
+
+async def comfy_entrypoint() -> FrameInterpolationExtension:
+    return FrameInterpolationExtension()

comfy_extras/nodes_hunyuan3d.py

@@ -637,7 +637,7 @@ class SaveGLB(IO.ComfyNode):
                     ],
                     tooltip="Mesh or 3D file to save",
                 ),
-                IO.String.Input("filename_prefix", default="mesh/ComfyUI"),
+                IO.String.Input("filename_prefix", default="3d/ComfyUI"),
             ],
             hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo]
         )

comfy_extras/nodes_lt.py

@@ -1,6 +1,7 @@
 import nodes
 import node_helpers
 import torch
+import torchaudio
 import comfy.model_management
 import comfy.model_sampling
 import comfy.samplers
@@ -711,7 +712,14 @@ class LTXVReferenceAudio(io.ComfyNode):
     @classmethod
     def execute(cls, model, positive, negative, reference_audio, audio_vae, identity_guidance_scale, start_percent, end_percent) -> io.NodeOutput:
         # Encode reference audio to latents and patchify
-        audio_latents = audio_vae.encode(reference_audio)
+        sample_rate = reference_audio["sample_rate"]
+        vae_sample_rate = getattr(audio_vae, "audio_sample_rate", 44100)
+        if vae_sample_rate != sample_rate:
+            waveform = torchaudio.functional.resample(reference_audio["waveform"], sample_rate, vae_sample_rate)
+        else:
+            waveform = reference_audio["waveform"]
+
+        audio_latents = audio_vae.encode(waveform.movedim(1, -1))
         b, c, t, f = audio_latents.shape
         ref_tokens = audio_latents.permute(0, 2, 1, 3).reshape(b, t, c * f)
         ref_audio = {"tokens": ref_tokens}

comfy_extras/nodes_mask.py

@@ -2,6 +2,7 @@ import numpy as np
 import scipy.ndimage
 import torch
 import comfy.utils
+import comfy.model_management
 import node_helpers
 from typing_extensions import override
 from comfy_api.latest import ComfyExtension, IO, UI
@@ -188,7 +189,7 @@ class SolidMask(IO.ComfyNode):
 
     @classmethod
     def execute(cls, value, width, height) -> IO.NodeOutput:
-        out = torch.full((1, height, width), value, dtype=torch.float32, device="cpu")
+        out = torch.full((1, height, width), value, dtype=torch.float32, device=comfy.model_management.intermediate_device())
         return IO.NodeOutput(out)
 
     solid = execute  # TODO: remove
@@ -262,6 +263,7 @@ class MaskComposite(IO.ComfyNode):
     def execute(cls, destination, source, x, y, operation) -> IO.NodeOutput:
         output = destination.reshape((-1, destination.shape[-2], destination.shape[-1])).clone()
         source = source.reshape((-1, source.shape[-2], source.shape[-1]))
+        source = source.to(output.device)
 
         left, top = (x, y,)
         right, bottom = (min(left + source.shape[-1], destination.shape[-1]), min(top + source.shape[-2], destination.shape[-2]))

comfy_extras/nodes_post_processing.py

@@ -666,12 +666,13 @@ class ColorTransfer(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="ColorTransfer",
+            display_name="Color Transfer",
             category="image/postprocessing",
             description="Match the colors of one image to another using various algorithms.",
             search_aliases=["color match", "color grading", "color correction", "match colors", "color transform", "mkl", "reinhard", "histogram"],
             inputs=[
                 io.Image.Input("image_target", tooltip="Image(s) to apply the color transform to."),
-                io.Image.Input("image_ref", optional=True, tooltip="Reference image(s) to match colors to. If not provided, processing is skipped"),
+                io.Image.Input("image_ref", tooltip="Reference image(s) to match colors to."),
                 io.Combo.Input("method", options=['reinhard_lab', 'mkl_lab', 'histogram'],),
                 io.DynamicCombo.Input("source_stats",
                     tooltip="per_frame: each frame matched to image_ref individually. uniform: pool stats across all source frames as baseline, match to image_ref. target_frame: use one chosen frame as the baseline for the transform to image_ref, applied uniformly to all frames (preserves relative differences)",

comfy_extras/nodes_preview_any.py

@@ -1,5 +1,6 @@
 import json
 from comfy.comfy_types.node_typing import IO
+import torch
 
 # Preview Any - original implement from
 # https://github.com/rgthree/rgthree-comfy/blob/main/py/display_any.py
@@ -19,6 +20,7 @@ class PreviewAny():
     SEARCH_ALIASES = ["show output", "inspect", "debug", "print value", "show text"]
 
     def main(self, source=None):
+        torch.set_printoptions(edgeitems=6)
         value = 'None'
         if isinstance(source, str):
             value = source
@@ -33,6 +35,7 @@ class PreviewAny():
                 except Exception:
                     value = 'source exists, but could not be serialized.'
 
+        torch.set_printoptions()
         return {"ui": {"text": (value,)}, "result": (value,)}
 
 NODE_CLASS_MAPPINGS = {

comfy_extras/nodes_primitive.py

@@ -49,7 +49,7 @@ class Int(io.ComfyNode):
             display_name="Int",
             category="utils/primitive",
             inputs=[
-                io.Int.Input("value", min=-sys.maxsize, max=sys.maxsize, control_after_generate=True),
+                io.Int.Input("value", min=-sys.maxsize, max=sys.maxsize, control_after_generate=io.ControlAfterGenerate.fixed),
             ],
             outputs=[io.Int.Output()],
         )