Fooocus Prompt Expansion (#329)

* add vae approx download

* files

* files

* files

* i

* i

* i

* i

* i

* i

* i

* i

* i

* i

.gitignore

@@ -2,6 +2,7 @@ __pycache__
 *.ckpt
 *.safetensors
 *.pth
+*.bin
 lena.png
 lena_result.png
 lena_test.py

fooocus_version.py

@@ -1 +1 @@
-version = '1.0.45'
+version = '1.0.50'

launch.py

@@ -6,7 +6,7 @@ import fooocus_version
 from modules.launch_util import is_installed, run, python, \
     run_pip, repo_dir, git_clone, requirements_met, script_path, dir_repos
 from modules.model_loader import load_file_from_url
-from modules.path import modelfile_path, lorafile_path
+from modules.path import modelfile_path, lorafile_path, vae_approx_path, fooocus_expansion_path
 
 REINSTALL_ALL = False
 
@@ -63,12 +63,26 @@ lora_filenames = [
      'https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_offset_example-lora_1.0.safetensors')
 ]
 
+vae_approx_filenames = [
+    ('taesdxl_decoder.pth',
+     'https://huggingface.co/lllyasviel/misc/resolve/main/taesdxl_decoder.pth')
+]
+
 
 def download_models():
     for file_name, url in model_filenames:
         load_file_from_url(url=url, model_dir=modelfile_path, file_name=file_name)
     for file_name, url in lora_filenames:
         load_file_from_url(url=url, model_dir=lorafile_path, file_name=file_name)
+    for file_name, url in vae_approx_filenames:
+        load_file_from_url(url=url, model_dir=vae_approx_path, file_name=file_name)
+
+    load_file_from_url(
+        url='https://huggingface.co/lllyasviel/misc/resolve/main/fooocus_expansion.bin',
+        model_dir=fooocus_expansion_path,
+        file_name='pytorch_model.bin'
+    )
+
     return
 
 

models/prompt_expansion/fooocus_expansion/config.json

@@ -0,0 +1,39 @@
+{
+  "_name_or_path": "gpt2",
+  "activation_function": "gelu_new",
+  "architectures": [
+    "GPT2LMHeadModel"
+  ],
+  "attn_pdrop": 0.1,
+  "bos_token_id": 50256,
+  "embd_pdrop": 0.1,
+  "eos_token_id": 50256,
+  "initializer_range": 0.02,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "gpt2",
+  "n_ctx": 1024,
+  "n_embd": 768,
+  "n_head": 12,
+  "n_inner": null,
+  "n_layer": 12,
+  "n_positions": 1024,
+  "reorder_and_upcast_attn": false,
+  "resid_pdrop": 0.1,
+  "scale_attn_by_inverse_layer_idx": false,
+  "scale_attn_weights": true,
+  "summary_activation": null,
+  "summary_first_dropout": 0.1,
+  "summary_proj_to_labels": true,
+  "summary_type": "cls_index",
+  "summary_use_proj": true,
+  "task_specific_params": {
+    "text-generation": {
+      "do_sample": true,
+      "max_length": 50
+    }
+  },
+  "torch_dtype": "float32",
+  "transformers_version": "4.23.0.dev0",
+  "use_cache": true,
+  "vocab_size": 50257
+}

models/prompt_expansion/fooocus_expansion/special_tokens_map.json

@@ -0,0 +1,5 @@
+{
+  "bos_token": "<|endoftext|>",
+  "eos_token": "<|endoftext|>",
+  "unk_token": "<|endoftext|>"
+}

models/prompt_expansion/fooocus_expansion/tokenizer_config.json

@@ -0,0 +1,10 @@
+{
+  "add_prefix_space": false,
+  "bos_token": "<|endoftext|>",
+  "eos_token": "<|endoftext|>",
+  "model_max_length": 1024,
+  "name_or_path": "gpt2",
+  "special_tokens_map_file": null,
+  "tokenizer_class": "GPT2Tokenizer",
+  "unk_token": "<|endoftext|>"
+}

modules/async_worker.py

@@ -15,7 +15,7 @@ def worker():
     import modules.path
     import modules.patch
 
-    from modules.sdxl_styles import apply_style, aspect_ratios
+    from modules.sdxl_styles import apply_style_negative, apply_style_positive, aspect_ratios
     from modules.private_logger import log
 
     try:
@@ -29,19 +29,69 @@ def worker():
 
     def handler(task):
         prompt, negative_prompt, style_selction, performance_selction, \
-        aspect_ratios_selction, image_number, image_seed, sharpness, base_model_name, refiner_model_name, \
+        aspect_ratios_selction, image_number, image_seed, sharpness, raw_mode, \
+        base_model_name, refiner_model_name, \
         l1, w1, l2, w2, l3, w3, l4, w4, l5, w5 = task
 
         loras = [(l1, w1), (l2, w2), (l3, w3), (l4, w4), (l5, w5)]
 
         modules.patch.sharpness = sharpness
 
+        outputs.append(['preview', (1, 'Initializing ...', None)])
+
+        seed = image_seed
+        max_seed = int(1024 * 1024 * 1024)
+        if not isinstance(seed, int):
+            seed = random.randint(1, max_seed)
+        if seed < 0:
+            seed = - seed
+        seed = seed % max_seed
+
+        outputs.append(['preview', (3, 'Load models ...', None)])
+
         pipeline.refresh_base_model(base_model_name)
         pipeline.refresh_refiner_model(refiner_model_name)
         pipeline.refresh_loras(loras)
-        pipeline.clean_prompt_cond_caches()
 
-        p_txt, n_txt = apply_style(style_selction, prompt, negative_prompt)
+        outputs.append(['preview', (5, 'Encoding negative text ...', None)])
+        n_txt = apply_style_negative(style_selction, negative_prompt)
+        n_cond = pipeline.process_prompt(n_txt)
+
+        tasks = []
+        if raw_mode:
+            outputs.append(['preview', (9, 'Encoding positive text ...', None)])
+            p_txt = apply_style_positive(style_selction, prompt)
+            p_cond = pipeline.process_prompt(p_txt)
+            for i in range(image_number):
+                tasks.append(dict(
+                    prompt=prompt,
+                    negative_prompt=negative_prompt,
+                    seed=seed + i,
+                    n_cond=n_cond,
+                    p_cond=p_cond,
+                    real_positive_prompt=p_txt,
+                    real_negative_prompt=n_txt
+                ))
+        else:
+            for i in range(image_number):
+                outputs.append(['preview', (9, f'Preparing positive text #{i + 1} ...', None)])
+                current_seed = seed + i
+
+                p_txt = pipeline.expand_txt(prompt, current_seed)
+                print(f'Expanded positive prompt: {p_txt}')
+
+                p_txt = apply_style_positive(style_selction, p_txt)
+                tasks.append(dict(
+                    prompt=prompt,
+                    negative_prompt=negative_prompt,
+                    seed=current_seed,
+                    n_cond=n_cond,
+                    real_positive_prompt=p_txt,
+                    real_negative_prompt=n_txt
+                ))
+            for i, t in enumerate(tasks):
+                outputs.append(['preview', (12, f'Encoding positive text #{i + 1} ...', None)])
+                t['p_cond'] = pipeline.process_prompt(t['real_positive_prompt'])
 
         if performance_selction == 'Speed':
             steps = 30
@@ -53,45 +103,47 @@ def worker():
         width, height = aspect_ratios[aspect_ratios_selction]
 
         results = []
-        seed = image_seed
-        max_seed = int(1024*1024*1024)
-
-        if not isinstance(seed, int):
-            seed = random.randint(1, max_seed)
-        if seed < 0:
-            seed = - seed
-        seed = seed % max_seed
-
         all_steps = steps * image_number
 
         def callback(step, x0, x, total_steps, y):
-            done_steps = i * steps + step
+            done_steps = current_task_id * steps + step
             outputs.append(['preview', (
-                int(100.0 * float(done_steps) / float(all_steps)),
+                int(15.0 + 85.0 * float(done_steps) / float(all_steps)),
                 f'Step {step}/{total_steps} in the {i}-th Sampling',
                 y)])
 
-        for i in range(image_number):
+        outputs.append(['preview', (13, 'Starting tasks ...', None)])
-            imgs = pipeline.process(p_txt, n_txt, steps, switch, width, height, seed, callback=callback)
+        for current_task_id, task in enumerate(tasks):
+            imgs = pipeline.process_diffusion(
+                positive_cond=task['p_cond'],
+                negative_cond=task['n_cond'],
+                steps=steps,
+                switch=switch,
+                width=width,
+                height=height,
+                image_seed=task['seed'],
+                callback=callback)
 
             for x in imgs:
                 d = [
-                    ('Prompt', prompt),
+                    ('Prompt', task['prompt']),
-                    ('Negative Prompt', negative_prompt),
+                    ('Negative Prompt', task['negative_prompt']),
+                    ('Real Positive Prompt', task['real_positive_prompt']),
+                    ('Real Negative Prompt', task['real_negative_prompt']),
+                    ('Raw Mode', str(raw_mode)),
                     ('Style', style_selction),
                     ('Performance', performance_selction),
                     ('Resolution', str((width, height))),
                     ('Sharpness', sharpness),
                     ('Base Model', base_model_name),
                     ('Refiner Model', refiner_model_name),
-                    ('Seed', seed)
+                    ('Seed', task['seed'])
                 ]
                 for n, w in loras:
                     if n != 'None':
                         d.append((f'LoRA [{n}] weight', w))
                 log(x, d)
 
-            seed += 1
             results += imgs
 
         outputs.append(['results', results])

modules/default_pipeline.py

@@ -5,6 +5,7 @@ import modules.path
 
 from comfy.model_base import SDXL, SDXLRefiner
 from modules.patch import cfg_patched
+from modules.expansion import FooocusExpansion
 
 
 xl_base: core.StableDiffusionModel = None
@@ -43,7 +44,6 @@ def refresh_base_model(name):
     xl_base_patched = xl_base
     xl_base_patched_hash = ''
     print(f'Base model loaded: {xl_base_hash}')
-
     return
 
 
@@ -103,27 +103,24 @@ refresh_base_model(modules.path.default_base_model_name)
 refresh_refiner_model(modules.path.default_refiner_model_name)
 refresh_loras([(modules.path.default_lora_name, 0.5), ('None', 0.5), ('None', 0.5), ('None', 0.5), ('None', 0.5)])
 
-positive_conditions_cache = None
+expansion_model = FooocusExpansion()
-negative_conditions_cache = None
-positive_conditions_refiner_cache = None
-negative_conditions_refiner_cache = None
 
 
-def clean_prompt_cond_caches():
+def expand_txt(*args, **kwargs):
-    global positive_conditions_cache, negative_conditions_cache, \
+    return expansion_model(*args, **kwargs)
-        positive_conditions_refiner_cache, negative_conditions_refiner_cache
+
-    positive_conditions_cache = None
+
-    negative_conditions_cache = None
+def process_prompt(text):
-    positive_conditions_refiner_cache = None
+    base_cond = core.encode_prompt_condition(clip=xl_base_patched.clip, prompt=text)
-    negative_conditions_refiner_cache = None
+    if xl_refiner is not None:
-    return
+        refiner_cond = core.encode_prompt_condition(clip=xl_refiner.clip, prompt=text)
+    else:
+        refiner_cond = None
+    return base_cond, refiner_cond
 
 
 @torch.no_grad()
-def process(positive_prompt, negative_prompt, steps, switch, width, height, image_seed, callback):
+def process_diffusion(positive_cond, negative_cond, steps, switch, width, height, image_seed, callback):
-    global positive_conditions_cache, negative_conditions_cache, \
-        positive_conditions_refiner_cache, negative_conditions_refiner_cache
-
     if xl_base is not None:
         xl_base.unet.model_options['sampler_cfg_function'] = cfg_patched
 
@@ -133,40 +130,27 @@ def process(positive_prompt, negative_prompt, steps, switch, width, height, imag
     if xl_refiner is not None:
         xl_refiner.unet.model_options['sampler_cfg_function'] = cfg_patched
 
-    positive_conditions = core.encode_prompt_condition(clip=xl_base_patched.clip, prompt=positive_prompt) if positive_conditions_cache is None else positive_conditions_cache
-    negative_conditions = core.encode_prompt_condition(clip=xl_base_patched.clip, prompt=negative_prompt) if negative_conditions_cache is None else negative_conditions_cache
-
-    positive_conditions_cache = positive_conditions
-    negative_conditions_cache = negative_conditions
-
     empty_latent = core.generate_empty_latent(width=width, height=height, batch_size=1)
 
     if xl_refiner is not None:
-        positive_conditions_refiner = core.encode_prompt_condition(clip=xl_refiner.clip, prompt=positive_prompt) if positive_conditions_refiner_cache is None else positive_conditions_refiner_cache
-        negative_conditions_refiner = core.encode_prompt_condition(clip=xl_refiner.clip, prompt=negative_prompt) if negative_conditions_refiner_cache is None else negative_conditions_refiner_cache
-
-        positive_conditions_refiner_cache = positive_conditions_refiner
-        negative_conditions_refiner_cache = negative_conditions_refiner
-
         sampled_latent = core.ksampler_with_refiner(
             model=xl_base_patched.unet,
-            positive=positive_conditions,
+            positive=positive_cond[0],
-            negative=negative_conditions,
+            negative=negative_cond[0],
             refiner=xl_refiner.unet,
-            refiner_positive=positive_conditions_refiner,
+            refiner_positive=positive_cond[1],
-            refiner_negative=negative_conditions_refiner,
+            refiner_negative=negative_cond[1],
             refiner_switch_step=switch,
             latent=empty_latent,
             steps=steps, start_step=0, last_step=steps, disable_noise=False, force_full_denoise=True,
             seed=image_seed,
             callback_function=callback
         )
-
     else:
         sampled_latent = core.ksampler(
             model=xl_base_patched.unet,
-            positive=positive_conditions,
+            positive=positive_cond[0],
-            negative=negative_conditions,
+            negative=negative_cond[0],
             latent=empty_latent,
             steps=steps, start_step=0, last_step=steps, disable_noise=False, force_full_denoise=True,
             seed=image_seed,
@@ -174,7 +158,5 @@ def process(positive_prompt, negative_prompt, steps, switch, width, height, imag
         )
 
     decoded_latent = core.decode_vae(vae=xl_base_patched.vae, latent_image=sampled_latent)
-
     images = core.image_to_numpy(decoded_latent)
-
     return images

modules/expansion.py

@@ -0,0 +1,23 @@
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline, set_seed
+from modules.path import fooocus_expansion_path
+
+
+class FooocusExpansion:
+    def __init__(self):
+        self.tokenizer = AutoTokenizer.from_pretrained(fooocus_expansion_path)
+        self.model = AutoModelForCausalLM.from_pretrained(fooocus_expansion_path)
+        self.pipe = pipeline('text-generation',
+                             model=self.model,
+                             tokenizer=self.tokenizer,
+                             device='cpu',
+                             torch_dtype=torch.float32)
+        print('Fooocus Expansion engine loaded.')
+
+    def __call__(self, prompt, seed):
+        prompt = str(prompt).rstrip('\n')
+        seed = int(seed)
+        set_seed(seed)
+        response = self.pipe(prompt, max_length=len(prompt) + 256)
+        result = response[0]['generated_text'].rstrip('\n')
+        return result

modules/path.py

@@ -2,8 +2,12 @@ import os
 
 modelfile_path = os.path.abspath(os.path.join(os.path.dirname(__file__), '../models/checkpoints/'))
 lorafile_path = os.path.abspath(os.path.join(os.path.dirname(__file__), '../models/loras/'))
+vae_approx_path = os.path.abspath(os.path.join(os.path.dirname(__file__), '../models/vae_approx/'))
 temp_outputs_path = os.path.abspath(os.path.join(os.path.dirname(__file__), '../outputs/'))
 
+fooocus_expansion_path = os.path.abspath(os.path.join(os.path.dirname(__file__),
+                                                      '../models/prompt_expansion/fooocus_expansion'))
+
 os.makedirs(temp_outputs_path, exist_ok=True)
 
 default_base_model_name = 'sd_xl_base_1.0_0.9vae.safetensors'

modules/private_logger.py

@@ -22,7 +22,7 @@ def log(img, dic):
         f.write(f"<p>{only_name}</p>\n")
         i = 0
         for k, v in dic:
-            if i < 2:
+            if i < 4:
                 f.write(f"<p>{k}: <b>{v}</b> </p>\n")
             else:
                 if i % 2 == 0:

modules/sdxl_styles.py

@@ -959,6 +959,14 @@ SD_XL_BASE_RATIOS = {
 aspect_ratios = {str(v[0]) + '×' + str(v[1]): v for k, v in SD_XL_BASE_RATIOS.items()}
 
 
-def apply_style(style, positive, negative):
+def apply_style_positive(style, txt):
     p, n = styles.get(style, default_style)
-    return p.replace('{prompt}', positive), n + ', ' + negative
+    return p.replace('{prompt}', txt)
+
+
+def apply_style_negative(style, txt):
+    p, n = styles.get(style, default_style)
+    if n == '':
+        return txt
+    else:
+        return n + ', ' + txt

readme.md

@@ -97,17 +97,16 @@ Coming soon ...
 
 Below things are already inside the software, and **users do not need to do anything about these**.
 
-~Note that some of these tricks are currently (2023 Aug 11) impossible to reproduce in Automatic1111's interface or ComfyUI's node system.~ (Update Aug 21: We are working on implementing some of these as webui extensions/features.)
+1. Prompt expansion and the "Raw mode" to turn it off (similar to Midjourney's hidden pre-processsing and "raw" mode, or the LeonardoAI's Prompt Magic).
-
+2. Native refiner swap inside one single k-sampler. The advantage is that now the refiner model can reuse the base model's momentum (or ODE's history parameters) collected from k-sampling to achieve more coherent sampling. In Automatic1111's high-res fix and ComfyUI's node system, the base model and refiner use two independent k-samplers, which means the momentum is largely wasted, and the sampling continuity is broken. Fooocus uses its own advanced k-diffusion sampling that ensures seamless, native, and continuous swap in a refiner setup. (Update Aug 13: Actually I discussed this with Automatic1111 several days ago and it seems that the “native refiner swap inside one single k-sampler” is [merged]( https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12371) into the dev branch of webui. Great!)
-1. Native refiner swap inside one single k-sampler. The advantage is that now the refiner model can reuse the base model's momentum (or ODE's history parameters) collected from k-sampling to achieve more coherent sampling. In Automatic1111's high-res fix and ComfyUI's node system, the base model and refiner use two independent k-samplers, which means the momentum is largely wasted, and the sampling continuity is broken. Fooocus uses its own advanced k-diffusion sampling that ensures seamless, native, and continuous swap in a refiner setup. (Update Aug 13: Actually I discussed this with Automatic1111 several days ago and it seems that the “native refiner swap inside one single k-sampler” is [merged]( https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12371) into the dev branch of webui. Great!)
+3. Negative ADM guidance. Because the highest resolution level of XL Base does not have cross attentions, the positive and negative signals for XL's highest resolution level cannot receive enough contrasts during the CFG sampling, causing the results look a bit plastic or overly smooth in certain cases. Fortunately, since the XL's highest resolution level is still conditioned on image aspect ratios (ADM), we can modify the adm on the positive/negative side to compensate for the lack of CFG contrast in the highest resolution level. (Update Aug 16, the IOS App [Drawing Things](https://apps.apple.com/us/app/draw-things-ai-generation/id6444050820) will support Negative ADM Guidance. Great!)
-2. Negative ADM guidance. Because the highest resolution level of XL Base does not have cross attentions, the positive and negative signals for XL's highest resolution level cannot receive enough contrasts during the CFG sampling, causing the results look a bit plastic or overly smooth in certain cases. Fortunately, since the XL's highest resolution level is still conditioned on image aspect ratios (ADM), we can modify the adm on the positive/negative side to compensate for the lack of CFG contrast in the highest resolution level. (Update Aug 16, the IOS App [Drawing Things](https://apps.apple.com/us/app/draw-things-ai-generation/id6444050820) will support Negative ADM Guidance. Great!)
+4. We implemented a carefully tuned variation of the Section 5.1 of ["Improving Sample Quality of Diffusion Models Using Self-Attention Guidance"](https://arxiv.org/pdf/2210.00939.pdf). The weight is set to very low, but this is Fooocus's final guarantee to make sure that the XL will never yield overly smooth or plastic appearance (examples [here](https://github.com/lllyasviel/Fooocus/discussions/117)). This can almostly eliminate all cases that XL still occasionally produce overly smooth results even with negative ADM guidance. (Update 2023 Aug 18, the Gaussian kernel of SAG is changed to an anisotropic kernel for better structure preservation and fewer artifacts.)
-3. We implemented a carefully tuned variation of the Section 5.1 of ["Improving Sample Quality of Diffusion Models Using Self-Attention Guidance"](https://arxiv.org/pdf/2210.00939.pdf). The weight is set to very low, but this is Fooocus's final guarantee to make sure that the XL will never yield overly smooth or plastic appearance (examples [here](https://github.com/lllyasviel/Fooocus/discussions/117)). This can almostly eliminate all cases that XL still occasionally produce overly smooth results even with negative ADM guidance. (Update 2023 Aug 18, the Gaussian kernel of SAG is changed to an anisotropic kernel for better structure preservation and fewer artifacts.)
+5. We modified the style templates a bit and added the "cinematic-default".
-4. We modified the style templates a bit and added the "cinematic-default".
+6. We tested the "sd_xl_offset_example-lora_1.0.safetensors" and it seems that when the lora weight is below 0.5, the results are always better than XL without lora.
-5. We tested the "sd_xl_offset_example-lora_1.0.safetensors" and it seems that when the lora weight is below 0.5, the results are always better than XL without lora.
+7. The parameters of samplers are carefully tuned.
-6. The parameters of samplers are carefully tuned.
+8. Because XL uses positional encoding for generation resolution, images generated by several fixed resolutions look a bit better than that from arbitrary resolutions (because the positional encoding is not very good at handling int numbers that are unseen during training). This suggests that the resolutions in UI may be hard coded for best results.
-7. Because XL uses positional encoding for generation resolution, images generated by several fixed resolutions look a bit better than that from arbitrary resolutions (because the positional encoding is not very good at handling int numbers that are unseen during training). This suggests that the resolutions in UI may be hard coded for best results.
+9. Separated prompts for two different text encoders seem unnecessary. Separated prompts for base model and refiner may work but the effects are random, and we refrain from implement this.
-8. Separated prompts for two different text encoders seem unnecessary. Separated prompts for base model and refiner may work but the effects are random, and we refrain from implement this.
+10. DPM family seems well-suited for XL, since XL sometimes generates overly smooth texture but DPM family sometimes generate overly dense detail in texture. Their joint effect looks neutral and appealing to human perception.
-9. DPM family seems well-suited for XL, since XL sometimes generates overly smooth texture but DPM family sometimes generate overly dense detail in texture. Their joint effect looks neutral and appealing to human perception.
 
 ## Advanced Features
 

update_log.md

@@ -1,3 +1,7 @@
+### 1.0.50
+
+* Prompt expansion and a "Raw mode" to turn it off (similar to Midjourney's "raw").
+
 ### 1.0.45
 
 * Reworked SAG, removed unnecessary patch

webui.py

@@ -56,10 +56,11 @@ with shared.gradio_root:
         with gr.Column(scale=0.5, visible=False) as right_col:
             with gr.Tab(label='Setting'):
                 performance_selction = gr.Radio(label='Performance', choices=['Speed', 'Quality'], value='Speed')
-                aspect_ratios_selction = gr.Radio(label='Aspect Ratios (width × height)', choices=list(aspect_ratios.keys()),
+                aspect_ratios_selction = gr.Radio(label='Aspect Ratios', choices=list(aspect_ratios.keys()),
-                                                  value='1152×896')
+                                                  value='1152×896', info='width × height')
                 image_number = gr.Slider(label='Image Number', minimum=1, maximum=32, step=1, value=2)
-                negative_prompt = gr.Textbox(label='Negative Prompt', show_label=True, placeholder="Type prompt here.")
+                negative_prompt = gr.Textbox(label='Negative Prompt', show_label=True, placeholder="Type prompt here.",
+                                             info='Describing objects that you do not want to see.')
                 seed_random = gr.Checkbox(label='Random', value=True)
                 image_seed = gr.Number(label='Seed', value=0, precision=0, visible=False)
 
@@ -75,8 +76,11 @@ with shared.gradio_root:
                 seed_random.change(random_checked, inputs=[seed_random], outputs=[image_seed])
 
             with gr.Tab(label='Style'):
-                style_selction = gr.Radio(show_label=False, container=True,
+                raw_mode_check = gr.Checkbox(label='Raw Mode', value=False,
-                                          choices=style_keys, value='cinematic-default')
+                                             info='Similar to Midjourney\'s \"raw\" mode.')
+                style_selction = gr.Radio(show_label=True, container=True,
+                                          choices=style_keys, value='cinematic-default', label='Image Style',
+                                          info='Similar to Midjourney\'s \"--style\".')
             with gr.Tab(label='Advanced'):
                 with gr.Row():
                     base_model = gr.Dropdown(label='SDXL Base Model', choices=modules.path.model_filenames, value=modules.path.default_base_model_name, show_label=True)
@@ -107,7 +111,7 @@ with shared.gradio_root:
         advanced_checkbox.change(lambda x: gr.update(visible=x), advanced_checkbox, right_col)
         ctrls = [
             prompt, negative_prompt, style_selction,
-            performance_selction, aspect_ratios_selction, image_number, image_seed, sharpness
+            performance_selction, aspect_ratios_selction, image_number, image_seed, sharpness, raw_mode_check
         ]
         ctrls += [base_model, refiner_model] + lora_ctrls
         run_button.click(fn=refresh_seed, inputs=[seed_random, image_seed], outputs=image_seed)\