app.py

import gradio as gr
import spaces
#import gradio.helpers
import torch
import os
import shutil
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 uuid
import random
from huggingface_hub import hf_hub_download

#gradio.helpers.CACHED_FOLDER = '/data/cache'

# OPTIONAL: Clear caches at startup to free space (comment out if you prefer not to)
hf_cache = os.path.expanduser("~/.cache/huggingface")
torch_cache = os.path.expanduser("~/.cache/torch")
if os.path.exists(hf_cache):
    shutil.rmtree(hf_cache)
if os.path.exists(torch_cache):
    shutil.rmtree(torch_cache)

# Load the pipeline with authentication token
pipe = StableVideoDiffusionPipeline.from_pretrained(
    "stabilityai/stable-video-diffusion-img2vid-xt",
    torch_dtype=torch.float16,
    variant="fp16",
    use_auth_token=os.getenv("HUGGINGFACE_TOKEN")  # Fetch the token from environment if set
)
pipe.to("cuda")
# Uncomment these lines only if you want experimental compilation (and if your environment supports it)
# pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
# pipe.vae = torch.compile(pipe.vae, mode="reduce-overhead", fullgraph=True)

max_64_bit_int = 2**63 - 1

def clean_outputs(output_folder: str, keep: int = 1):
    """
    Remove old video files to prevent using all disk space.
    Keeps the most recent <keep> files.
    """
    files = sorted(glob(os.path.join(output_folder, "*.mp4")), key=os.path.getmtime)
    if len(files) > keep:
        for old_file in files[:-keep]:
            os.remove(old_file)

@spaces.GPU(duration=250)
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",
    progress=gr.Progress(track_tqdm=True)
):
    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")

    # Reduce num_frames from 25 to 10 to consume less space
    frames = pipe(
        image, 
        decode_chunk_size=decoding_t, 
        generator=generator, 
        motion_bucket_id=motion_bucket_id, 
        noise_aug_strength=0.1, 
        num_frames=10  # reduced from 25
    ).frames[0]
    
    export_to_video(frames, video_path, fps=fps_id)
    torch.manual_seed(seed)

    # Clean up old videos to prevent filling disk
    clean_outputs(output_folder, keep=2)
    
    return video_path, seed

def resize_image(image, output_size=(1024, 576)):
    # Calculate aspect ratios
    target_aspect = output_size[0] / output_size[1]
    image_aspect = image.width / image.height

    # Resize then crop if the original image is larger
    if image_aspect > target_aspect:
        new_height = output_size[1]
        new_width = int(new_height * image_aspect)
        resized_image = image.resize((new_width, new_height), Image.LANCZOS)
        left = (new_width - output_size[0]) / 2
        top = 0
        right = (new_width + output_size[0]) / 2
        bottom = output_size[1]
    else:
        new_width = output_size[0]
        new_height = int(new_width / image_aspect)
        resized_image = image.resize((new_width, new_height), Image.LANCZOS)
        left = 0
        top = (new_height - output_size[1]) / 2
        right = output_size[0]
        bottom = (new_height + output_size[1]) / 2

    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 (`10 frames` at `6 fps`). This demo uses [🧨 diffusers](https://huggingface.co/docs/diffusers/main/en/using-diffusers/svd) for low VRAM usage.
    ''')
    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 num_frames/fps", value=6, minimum=5, maximum=30)

    # Resize on upload
    image.upload(fn=resize_image, inputs=image, outputs=image, queue=False)

    # Generate with sample() function
    generate_btn.click(
        fn=sample,
        inputs=[image, seed, randomize_seed, motion_bucket_id, fps_id],
        outputs=[video, seed],
        api_name="video"
    )

# REMOVED: gr.Examples(...) that referenced local images like "blink_meme.png" which don't exist

if __name__ == "__main__":
    # remove `share=True` because it is not supported on Hugging Face Spaces
    demo.launch(show_api=False)