app.py

import gradio as gr
import torch
import os
from glob import glob
from pathlib import Path
from typing import Optional

from diffusers import StableVideoDiffusionPipeline
from diffusers.utils import export_to_video
from PIL import Image

import random
import spaces

pipe = StableVideoDiffusionPipeline.from_pretrained(
    "vdo/stable-video-diffusion-img2vid-xt-1-1", torch_dtype=torch.float16, variant="fp16"
)
pipe.to("cuda")

max_64_bit_int = 2**63 - 1

@spaces.GPU(duration=120)
def sample(
    image: Image,
    seed: Optional[int] = 42,
    randomize_seed: bool = True,
    motion_bucket_id: int = 127,
    fps_id: int = 6,
    noise_aug_strength: float = 0.1,
    decoding_t: int = 3,
    frame_format: str = "webp",
    version: str = "svd_xt",
    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=noise_aug_strength, num_frames=25).frames[0]
    export_to_video(frames, video_path, fps=fps_id)
    
    return video_path, gr.update(label="Generated frames in *." + frame_format + " format", format = frame_format, value = frames), seed

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

    # Do not touch the image if the size is good
    if image.width == output_size[0] and image.height == output_size[1]:
        return image

    # Resize 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

    # Crop the image
    cropped_image = resized_image.crop((left, top, right, bottom))
    return cropped_image

with gr.Blocks() as demo:
  gr.Markdown('''# Community demo for Stable Video Diffusion - Img2Vid - XT ([model](https://huggingface.co/stabilityai/stable-video-diffusion-img2vid-xt), [paper](https://stability.ai/research/stable-video-diffusion-scaling-latent-video-diffusion-models-to-large-datasets), [stability's ui waitlist](https://stability.ai/contact))
#### Research release ([_non-commercial_](https://huggingface.co/stabilityai/stable-video-diffusion-img2vid-xt/blob/main/LICENSE)): generate `4s` vid from a single image at (`25 frames` at `6 fps`). this demo uses [🧨 diffusers for low VRAM and fast generation](https://huggingface.co/docs/diffusers/main/en/using-diffusers/svd).
  ''')
  with gr.Row():
      with gr.Column():
          image = gr.Image(label="Upload your image", type="pil")
          with gr.Accordion("Advanced options", open=False):
              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)
              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)
              noise_aug_strength = gr.Slider(label="Noise strength", info="The noise to add", value=0.1, minimum=0, maximum=1, step=0.1)
              decoding_t = gr.Slider(label="Decoding", info="Number of frames decoded at a time; this eats more VRAM; reduce if necessary", value=3, minimum=1, maximum=5, step=1)
              frame_format = gr.Radio([["*.png", "png"], ["*.webp", "webp"], ["*.jpeg", "jpeg"], ["*.gif", "gif"], ["*.bmp", "bmp"]], label="Image format for result", info="File extention", value="webp", interactive=True)
              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)

          generate_btn = gr.Button(value="Animate", variant="primary")

      with gr.Column():
          video = gr.Video(label="Generated video")
          gallery = gr.Gallery(label="Generated frames")
      
  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, noise_aug_strength, decoding_t, frame_format], outputs=[video, gallery, seed], api_name="video")
    
  gr.Examples(
    examples=[
        ["Examples/Fire.webp", 25, 127, 0.1, 3, "png", 42, True],
        ["Examples/Town.jpeg", 25, 127, 0.1, 3, "png", 42, True],
        ["Examples/Water.png", 25, 127, 0.1, 3, "png", 42, True]
    ],
    inputs=[image, fps_id, motion_bucket_id, noise_aug_strength, decoding_t, frame_format, seed, randomize_seed],
    outputs=[video, gallery, seed],
    fn=sample,
    run_on_click=True,
    cache_examples=False,
  )

if __name__ == "__main__":
    demo.launch(share=True, show_api=False)