Update app.py

app.py

@@ -1,166 +1,174 @@
+import random
+import spaces
+
 import gradio as gr
 import numpy as np
-import random
 import torch
-import spaces
-from PIL import Image
+from diffusers import (
+    StableDiffusion3Pipeline,
+    SD3Transformer2DModel,
+    FlashFlowMatchEulerDiscreteScheduler,
+    AutoencoderTiny,
+)
+from peft import PeftModel
 import os
+from huggingface_hub import snapshot_download
 
-from models.transformer_sd3 import SD3Transformer2DModel
-from pipeline_stable_diffusion_3_ipa import StableDiffusion3Pipeline
-
-from transformers import AutoProcessor, SiglipVisionModel
-from huggingface_hub import hf_hub_download
+huggingface_token = os.getenv("HUGGINFACE_TOKEN")
 
+model_path = snapshot_download(
+    repo_id="stabilityai/stable-diffusion-3-medium-diffusers",
+    repo_type="model",
+    ignore_patterns=["*.md", "*..gitattributes"],
+    local_dir="stable-diffusion-3-medium",
+    token=huggingface_token,  # type a new token-id.
+)
+import spaces
 
-# Constants
-MAX_SEED = np.iinfo(np.int32).max
-MAX_IMAGE_SIZE = 1024
-DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
-
-model_path = 'stabilityai/stable-diffusion-3.5-large'
-image_encoder_path = "google/siglip-so400m-patch14-384"
-ipadapter_path = hf_hub_download(repo_id="InstantX/SD3.5-Large-IP-Adapter", filename="ip-adapter.bin")
+device = "cuda" if torch.cuda.is_available() else "cpu"
 
 transformer = SD3Transformer2DModel.from_pretrained(
-    model_path, 
-    subfolder="transformer", 
-    torch_dtype=torch.bfloat16
+    model_path,
+    subfolder="transformer",
+    torch_dtype=torch.float16,
 )
+transformer = PeftModel.from_pretrained(transformer, "jasperai/flash-sd3")
+
+
+if torch.cuda.is_available():
+    torch.cuda.max_memory_allocated(device=device)
+    pipe = StableDiffusion3Pipeline.from_pretrained(
+        model_path,
+        transformer=transformer,
+        torch_dtype=torch.float16,
+        text_encoder_3=None,
+        tokenizer_3=None,
+    )
+    pipe.vae = AutoencoderTiny.from_pretrained(
+        "madebyollin/taesd3", torch_dtype=torch.float16
+    )
+
+    # pipe.vae.decoder.layers = torch.compile(
+    #     pipe.vae.decoder.layers,
+    #     fullgraph=True,
+    #     dynamic=False,
+    #     mode="max-autotune-no-cudagraphs",
+    # )
+    pipe.vae.config.shift_factor = 0.0
+
+    pipe = pipe.to(device)
+else:
+    pipe = StableDiffusion3Pipeline.from_pretrained(
+        model_path,
+        transformer=transformer,
+        torch_dtype=torch.float16,
+        text_encoder_3=None,
+        tokenizer_3=None,
+    )
+    pipe = pipe.to(device)
 
-pipe = StableDiffusion3Pipeline.from_pretrained(
-    model_path, 
-    transformer=transformer, 
-    torch_dtype=torch.bfloat16
-).to("cuda")
 
-pipe.init_ipadapter(
-    ip_adapter_path=ipadapter_path, 
-    image_encoder_path=image_encoder_path, 
-    nb_token=64, 
+pipe.scheduler = FlashFlowMatchEulerDiscreteScheduler.from_pretrained(
+    model_path,
+    subfolder="scheduler",
 )
+pipe.set_progress_bar_config(disable=True)
+
+
+MAX_SEED = np.iinfo(np.int32).max
+MAX_IMAGE_SIZE = 1024
+NUM_INFERENCE_STEPS = 4
 
-def resize_img(image, max_size=1024):
-    width, height = image.size
-    scaling_factor = min(max_size / width, max_size / height)
-    new_width = int(width * scaling_factor)
-    new_height = int(height * scaling_factor)
-    return image.resize((new_width, new_height), Image.LANCZOS)
 
 @spaces.GPU
-def process_image(
-    image,
-    prompt,
-    scale,
-    seed,
-    randomize_seed,
-    width,
-    height,
-    progress=gr.Progress(track_tqdm=True),
-):
-    #pipe.to("cuda")
+def infer(prompt, seed, randomize_seed):
     if randomize_seed:
         seed = random.randint(0, MAX_SEED)
-    
-    if image is None:
-        return None, seed
-    
-    # Convert to PIL Image if needed
-    if not isinstance(image, Image.Image):
-        image = Image.fromarray(image)
-    
-    # Resize image
-    image = resize_img(image)
-    
-    # Generate the image
-    result = pipe(
-        clip_image=image,
+
+    generator = torch.Generator().manual_seed(seed)
+
+    image = pipe(
         prompt=prompt,
-        ipadapter_scale=scale,
-        width=width,
-        height=height,
-        generator=torch.Generator().manual_seed(seed)
+        guidance_scale=0,
+        num_inference_steps=NUM_INFERENCE_STEPS,
+        generator=generator,
     ).images[0]
-    
-    return result, seed
 
-# UI CSS
+    return image
+
+
+examples = [
+    "The image showcases a freshly baked bread, possibly focaccia, with rosemary sprigs and red pepper flakes sprinkled on top. It's sliced and placed on a wire cooling rack, with a bowl of mixed peppercorns beside it.",
+    'a 3D render of a wizard raccoon holding a sign saying "SD3" with a magic wand.',
+    "A panda reading a book in a lush forest.",
+    "A raccoon trapped inside a glass jar full of colorful candies, the background is steamy with vivid colors",
+    "Pirate ship sailing on a sea with the milky way galaxy in the sky and purple glow lights",
+    "a cute cartoon fluffy rabbit pilot walking on a military aircraft carrier, 8k, cinematic",
+    "A 3d render of a futuristic city with a giant robot in the middle full of neon lights, pink and blue colors",
+    "A close up of an old elderly man with green eyes looking straight at the camera",
+    "photo of a huge red cat with green eyes sitting on a cloud in the sky, looking at the camera",
+]
+
 css = """
 #col-container {
     margin: 0 auto;
-    max-width: 960px;
+    max-width: 712px;
 }
 """
 
-# Create the Gradio interface
 with gr.Blocks(css=css) as demo:
     with gr.Column(elem_id="col-container"):
-        gr.Markdown("# InstantX's SD3.5 IP Adapter")
-        
+        gr.Markdown(
+            f"""
+        # ⚡ Flash Diffusion: FlashSD3 + TAESD3 ⚡️
+        [Flash Diffusion](https://gojasper.github.io/flash-diffusion-project/) with [Tiny AutoEncoder for Stable Diffusion 3](https://huggingface.co/madebyollin/taesd3)
+        """
+        )
+
         with gr.Row():
-            with gr.Column():
-                input_image = gr.Image(
-                    label="Input Image",
-                    type="pil"
-                )
-                scale = gr.Slider(
-                    label="Image Scale",
-                    minimum=0.0,
-                    maximum=1.0,
-                    step=0.1,
-                    value=0.7,
-                )
-                prompt = gr.Text(
-                    label="Prompt",
-                    max_lines=1,
-                    placeholder="Enter your prompt",
-                )
-                run_button = gr.Button("Generate", variant="primary")
-            
-            with gr.Column():
-                result = gr.Image(label="Result")
-        
+            prompt = gr.Text(
+                label="Prompt",
+                show_label=False,
+                max_lines=1,
+                placeholder="Enter your prompt",
+                container=False,
+            )
+
+            run_button = gr.Button("Run", scale=0)
+
+        result = gr.Image(label="Result", show_label=False)
+
         with gr.Accordion("Advanced Settings", open=False):
             seed = gr.Slider(
                 label="Seed",
                 minimum=0,
                 maximum=MAX_SEED,
                 step=1,
-                value=42,
+                value=0,
             )
-            
+
             randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
-            
-            with gr.Row():
-                width = gr.Slider(
-                    label="Width",
-                    minimum=256,
-                    maximum=MAX_IMAGE_SIZE,
-                    step=32,
-                    value=1024,
-                )
-                
-                height = gr.Slider(
-                    label="Height",
-                    minimum=256,
-                    maximum=MAX_IMAGE_SIZE,
-                    step=32,
-                    value=1024,
-                )
-    
-    run_button.click(
-        fn=process_image,
-        inputs=[
-            input_image,
-            prompt,
-            scale,
-            seed,
-            randomize_seed,
-            width,
-            height,
+
+        examples = gr.Examples(examples=examples, inputs=[prompt], cache_examples=False)
+
+        gr.Markdown("**Disclaimer:**")
+        gr.Markdown(
+            "This demo is only for research purpose. Jasper cannot be held responsible for the generation of NSFW (Not Safe For Work) content through the use of this demo. Users are solely responsible for any content they create, and it is their obligation to ensure that it adheres to appropriate and ethical standards. Jasper provides the tools, but the responsibility for their use lies with the individual user."
+        )
+    gr.on(
+        [
+            run_button.click,
+            seed.change,
+            randomize_seed.change,
+            prompt.submit,
+            prompt.change,
         ],
-        outputs=[result, seed],
+        fn=infer,
+        inputs=[prompt, seed, randomize_seed],
+        outputs=[result],
+        show_progress="minimal",
+        show_api=True,
+        trigger_mode="always_last",
     )
 
-if __name__ == "__main__":
-    demo.launch()
+demo.queue().launch(show_api=True)