Make the space work (#1)

- Make the space work (5d75032e00f28fc16e7248fbf02107ca507ecc77)


Co-authored-by: Fabrice TIERCELIN <Fabrice-TIERCELIN@users.noreply.huggingface.co>

app.py

@@ -1,36 +1,120 @@
 import gradio as gr
-from diffusers import DiffusionPipeline
 import torch
+import os
+from glob import glob
+from pathlib import Path
+from typing import Optional
+
+from diffusers import StableVideoDiffusionPipeline
+from diffusers.utils import load_image, export_to_video
 from PIL import Image
-import spaces
-
-# Load the pre-trained pipeline
-pipeline = DiffusionPipeline.from_pretrained("stabilityai/stable-video-diffusion-img2vid-xt")
-
-# Define the Gradio interface
-interface = gr.Interface(
-    fn=lambda img: generate_video(img),
-    inputs=gr.Image(type="pil"),
-    outputs=gr.Video(),
-    title="Stable Video Diffusion",
-    description="Upload an image to generate a video",
-    theme="soft"
+
+import uuid
+import random
+from huggingface_hub import hf_hub_download
+
+pipe = StableVideoDiffusionPipeline.from_pretrained(
+    "multimodalart/stable-video-diffusion", torch_dtype=torch.float16, variant="fp16"
 )
+pipe.to("cuda")
+pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
+
+max_64_bit_int = 2**63 - 1
+
+def sample(
+    image: Image,
+    seed: Optional[int] = 42,
+    randomize_seed: bool = True,
+    motion_bucket_id: int = 127,
+    fps_id: int = 6,
+    version: str = "svd_xt",
+    cond_aug: float = 0.02,
+    decoding_t: int = 3,  # Number of frames decoded at a time! This eats most VRAM. Reduce if necessary.
+    device: str = "cuda",
+    output_folder: str = "outputs",
+):
+    if image.mode == "RGBA":
+        image = image.convert("RGB")
+        
+    if(randomize_seed):
+        seed = random.randint(0, max_64_bit_int)
+    generator = torch.manual_seed(seed)
+    
+    os.makedirs(output_folder, exist_ok=True)
+    base_count = len(glob(os.path.join(output_folder, "*.mp4")))
+    video_path = os.path.join(output_folder, f"{base_count:06d}.mp4")
+
+    frames = pipe(image, decode_chunk_size=decoding_t, generator=generator, motion_bucket_id=motion_bucket_id, noise_aug_strength=0.1, num_frames=25).frames[0]
+    export_to_video(frames, video_path, fps=fps_id)
+    torch.manual_seed(seed)
+    
+    return video_path, seed
 
-@spaces.GPU(duration=250)
-def generate_video(image):
-  """
-  Generates a video from an input image using the pipeline.
+def resize_image(image, output_size=(1024, 576)):
+    # Calculate aspect ratios
+    target_aspect = output_size[0] / output_size[1]  # Aspect ratio of the desired size
+    image_aspect = image.width / image.height  # Aspect ratio of the original image
 
-  Args:
-      image: A PIL Image object representing the input image.
+    # Resize then crop if the original image is larger
+    if image_aspect > target_aspect:
+        # Resize the image to match the target height, maintaining aspect ratio
+        new_height = output_size[1]
+        new_width = int(new_height * image_aspect)
+        resized_image = image.resize((new_width, new_height), Image.LANCZOS)
+        # Calculate coordinates for cropping
+        left = (new_width - output_size[0]) / 2
+        top = 0
+        right = (new_width + output_size[0]) / 2
+        bottom = output_size[1]
+    else:
+        # Resize the image to match the target width, maintaining aspect ratio
+        new_width = output_size[0]
+        new_height = int(new_width / image_aspect)
+        resized_image = image.resize((new_width, new_height), Image.LANCZOS)
+        # Calculate coordinates for cropping
+        left = 0
+        top = (new_height - output_size[1]) / 2
+        right = output_size[0]
+        bottom = (new_height + output_size[1]) / 2
 
-  Returns:
-      A list of PIL Images representing the video frames.
-  """
-  video_frames = pipeline(image=image, num_inference_steps=20).images
+    # Crop the image
+    cropped_image = resized_image.crop((left, top, right, bottom))
+    return cropped_image
 
-  return video_frames
+with gr.Blocks() as demo:
+  gr.Markdown('''# Stable Video Diffusion
+  ''')
+  with gr.Row():
+    with gr.Column():
+        image = gr.Image(label="Upload your image", type="pil")
+        generate_btn = gr.Button("Generate")
+    video = gr.Video()
+  with gr.Accordion("Advanced options", open=False):
+      seed = gr.Slider(label="Seed", value=42, randomize=True, minimum=0, maximum=max_64_bit_int, step=1)
+      randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
+      motion_bucket_id = gr.Slider(label="Motion bucket id", info="Controls how much motion to add/remove from the image", value=127, minimum=1, maximum=255)
+      fps_id = gr.Slider(label="Frames per second", info="The length of your video in seconds will be 25/fps", value=6, minimum=5, maximum=30)
+      
+  image.upload(fn=resize_image, inputs=image, outputs=image, queue=False)
+  generate_btn.click(fn=sample, inputs=[image, seed, randomize_seed, motion_bucket_id, fps_id], outputs=[video, seed], api_name="video")
+  gr.Examples(
+    examples=[
+        "images/blink_meme.png",
+        "images/confused2_meme.png",
+        "images/disaster_meme.png",
+        "images/distracted_meme.png",
+        "images/hide_meme.png",
+        "images/nazare_meme.png",
+        "images/success_meme.png",
+        "images/willy_meme.png",
+        "images/wink_meme.png"
+    ],
+    inputs=image,
+    outputs=[video, seed],
+    fn=sample,
+    cache_examples=True,
+  )
 
-# Launch the Gradio app
-interface.launch()
+if __name__ == "__main__":
+    demo.queue(max_size=20, api_open=False)
+    demo.launch(share=True, show_api=False)