improve anime

Improve Fooocus Anime a bit by using better SD1.5 refining formulation.

fooocus_extras/vae_interpose.py

@@ -69,7 +69,7 @@ vae_approx_filename = os.path.join(vae_approx_path, 'xl-to-v1_interposer-v3.1.sa
 def parse(x):
     global vae_approx_model
 
-    x_origin = x['samples'].clone()
+    x_origin = x.clone()
 
     if vae_approx_model is None:
         model = Interposer()
@@ -89,6 +89,5 @@ def parse(x):
     fcbh.model_management.load_model_gpu(vae_approx_model)
 
     x = x_origin.to(device=vae_approx_model.load_device, dtype=vae_approx_model.dtype)
-    x = vae_approx_model.model(x)
+    x = vae_approx_model.model(x).to(x_origin)
-
+    return x
-    return {'samples': x.to(x_origin)}

fooocus_version.py

@@ -1 +1 @@
-version = '2.1.722'
+version = '2.1.723'

modules/core.py

@@ -218,19 +218,21 @@ def get_previewer(model):
 def ksampler(model, positive, negative, latent, seed=None, steps=30, cfg=7.0, sampler_name='dpmpp_2m_sde_gpu',
              scheduler='karras', denoise=1.0, disable_noise=False, start_step=None, last_step=None,
              force_full_denoise=False, callback_function=None, refiner=None, refiner_switch=-1,
-             previewer_start=None, previewer_end=None, sigmas=None, noise=None):
+             previewer_start=None, previewer_end=None, sigmas=None, extra_noise=0.0):
 
     if sigmas is not None:
         sigmas = sigmas.clone().to(fcbh.model_management.get_torch_device())
 
     latent_image = latent["samples"]
 
-    if noise is None:
+    if disable_noise:
-        if disable_noise:
+        noise = torch.zeros(latent_image.size(), dtype=latent_image.dtype, layout=latent_image.layout, device="cpu")
-            noise = torch.zeros(latent_image.size(), dtype=latent_image.dtype, layout=latent_image.layout, device="cpu")
+    else:
-        else:
+        batch_inds = latent["batch_index"] if "batch_index" in latent else None
-            batch_inds = latent["batch_index"] if "batch_index" in latent else None
+        noise = fcbh.sample.prepare_noise(latent_image, seed, batch_inds)
-            noise = fcbh.sample.prepare_noise(latent_image, seed, batch_inds)
+
+    if extra_noise > 0.0:
+        noise = noise * (1.0 + extra_noise)
 
     noise_mask = None
     if "noise_mask" in latent:

modules/default_pipeline.py

@@ -6,7 +6,7 @@ import modules.path
 import fcbh.model_management
 import fcbh.latent_formats
 import modules.inpaint_worker
-import modules.sample_hijack as sample_hijack
+import fooocus_extras.vae_interpose as vae_interpose
 
 from fcbh.model_base import SDXL, SDXLRefiner
 from modules.expansion import FooocusExpansion
@@ -270,22 +270,14 @@ refresh_everything(
 
 @torch.no_grad()
 @torch.inference_mode()
-def vae_parse(x, tiled=False, use_interpose=True):
+def vae_parse(latent, k=1.0):
-    if final_vae is None or final_refiner_vae is None:
+    if final_refiner_vae is None:
-        return x
+        result = latent["samples"]
-
-    if use_interpose:
-        print('VAE interposing ...')
-        import fooocus_extras.vae_interpose
-        x = fooocus_extras.vae_interpose.parse(x)
-        print('VAE interposed ...')
     else:
-        print('VAE parsing ...')
+        result = vae_interpose.parse(latent["samples"])
-        x = core.decode_vae(vae=final_vae, latent_image=x, tiled=tiled)
+    if k != 1.0:
-        x = core.encode_vae(vae=final_refiner_vae, pixels=x, tiled=tiled)
+        result = result * k
-        print('VAE parsed ...')
+    return {'samples': result}
-
-    return x
 
 
 @torch.no_grad()
@@ -444,8 +436,7 @@ def process_diffusion(positive_cond, negative_cond, steps, switch, width, height
         if modules.inpaint_worker.current_task is not None:
             modules.inpaint_worker.current_task.unswap()
 
-        sample_hijack.history_record = []
+        sampled_latent = core.ksampler(
-        core.ksampler(
             model=final_unet,
             positive=positive_cond,
             negative=negative_cond,
@@ -467,34 +458,20 @@ def process_diffusion(positive_cond, negative_cond, steps, switch, width, height
             target_model = final_unet
             print('Use base model to refine itself - this may because of developer mode.')
 
+        # Fooocus' vae parameters
+        k_data = 1.05
+        k_noise = 0.15
+        k_sigmas = 1.4
+
+        sampled_latent = vae_parse(sampled_latent, k=k_data)
+
         sigmas = calculate_sigmas(sampler=sampler_name,
                                   scheduler=scheduler_name,
                                   model=target_model.model,
                                   steps=steps,
-                                  denoise=denoise)[switch:]
+                                  denoise=denoise)[switch:] * k_sigmas
-        k1 = target_model.model.latent_format.scale_factor
-        k2 = final_unet.model.latent_format.scale_factor
-        k_sigmas = float(k1) / float(k2)
-        sigmas = sigmas * k_sigmas
         len_sigmas = len(sigmas) - 1
 
-        last_step, last_clean_latent, last_noisy_latent = sample_hijack.history_record[-1]
-        last_clean_latent = final_unet.model.process_latent_out(last_clean_latent.cpu().to(torch.float32))
-        last_noisy_latent = final_unet.model.process_latent_out(last_noisy_latent.cpu().to(torch.float32))
-        last_noise = last_noisy_latent - last_clean_latent
-        last_noise = last_noise / last_noise.std()
-
-        noise_mean = torch.mean(last_noise, dim=1, keepdim=True).repeat(1, 4, 1, 1) / k_sigmas
-
-        refiner_noise = torch.normal(
-            mean=noise_mean,
-            std=torch.ones_like(noise_mean),
-            generator=torch.manual_seed(image_seed+1)  # Avoid artifacts
-        ).to(last_noise)
-
-        sampled_latent = {'samples': last_clean_latent}
-        sampled_latent = vae_parse(sampled_latent)
-
         if modules.inpaint_worker.current_task is not None:
             modules.inpaint_worker.current_task.swap()
 
@@ -504,7 +481,7 @@ def process_diffusion(positive_cond, negative_cond, steps, switch, width, height
             negative=clip_separate(negative_cond, target_model=target_model.model, target_clip=final_clip),
             latent=sampled_latent,
             steps=len_sigmas, start_step=0, last_step=len_sigmas, disable_noise=False, force_full_denoise=True,
-            seed=image_seed+2,  # Avoid artifacts
+            seed=image_seed,
             denoise=denoise,
             callback_function=callback,
             cfg=cfg_scale,
@@ -513,7 +490,7 @@ def process_diffusion(positive_cond, negative_cond, steps, switch, width, height
             previewer_start=switch,
             previewer_end=steps,
             sigmas=sigmas,
-            noise=refiner_noise
+            extra_noise=k_noise
         )
 
         target_model = final_refiner_vae
@@ -522,5 +499,4 @@ def process_diffusion(positive_cond, negative_cond, steps, switch, width, height
         decoded_latent = core.decode_vae(vae=target_model, latent_image=sampled_latent, tiled=tiled)
 
     images = core.pytorch_to_numpy(decoded_latent)
-    sample_hijack.history_record = None
     return images

modules/sample_hijack.py

@@ -11,7 +11,6 @@ from fcbh.samplers import resolve_areas_and_cond_masks, wrap_model, calculate_st
 
 current_refiner = None
 refiner_switch_step = -1
-history_record = None
 
 
 @torch.no_grad()
@@ -118,9 +117,6 @@ def sample_hacked(model, noise, positive, negative, cfg, device, sampler, sigmas
         return
 
     def callback_wrap(step, x0, x, total_steps):
-        global history_record
-        if isinstance(history_record, list):
-            history_record.append((step, x0, x))
         if step == refiner_switch_step and current_refiner is not None:
             refiner_switch()
         if callback is not None:

update_log.md

@@ -1,3 +1,7 @@
+# 2.1.723
+
+* Improve Fooocus Anime a bit by using better SD1.5 refining formulation.
+
 # 2.1.722
 
 * Now it is possible to translate 100% all texts in the UI.