app.py

import spaces
import gradio as gr
from inference_utils import inference


@spaces.GPU(duration=120)
def send_to_model(source_video, prompt, neg_prompt, guidance_scale, video_length, old_qk):
    return inference(prompt=prompt, neg_prompt=neg_prompt, guidance_scale=guidance_scale, video_length=video_length, video_path=source_video, old_qk=old_qk)


if __name__ == "__main__":
    with gr.Blocks() as demo:
        gr.HTML(
            """
            <h1 style="text-align: center; font-size: 32px; font-family: 'Times New Roman', Times, serif;">
                FLATTEN: optical FLow-guided ATTENtion for consistent text-to-video editing
            </h1>
            <p style="text-align: center; font-size: 20px; font-family: 'Times New Roman', Times, serif;">
                <a style="text-align: center; display:inline-block"
                    href="https://flatten-video-editing.github.io/">
                    <img src="https://huggingface.co/datasets/huggingface/badges/raw/main/paper-page-sm.svg#center"
                    alt="Paper Page">
                </a>
                <a style="text-align: center; display:inline-block" href="https://huggingface.co/spaces/sky24h/FLATTEN-unofficial?duplicate=true">
                    <img src="https://huggingface.co/datasets/huggingface/badges/raw/main/duplicate-this-space-sm.svg#center" alt="Duplicate Space">
                </a>
            </p>
            """
        )
        gr.Interface(
            fn=send_to_model,
            inputs=[
                gr.Video(value=None, label="Source Image"),
                gr.Textbox(value="", label="Prompt"),
                gr.Textbox(value="", label="Negative Prompt"),
                gr.Slider(
                    value   = 15,
                    minimum = 10,
                    maximum = 30,
                    step    = 1,
                    label   = "guidance_scale",
                    info    = "The scale of the guidance field.",
                ),
                gr.Textbox(value=16, label="Video Length", info="The length of the video, must be less than 16 frames in the online demo to avoid timeout. However, you can run the model locally to process longer videos."),
                gr.Dropdown(value=0, choices=[0, 1], label="Choose Option", info="Select 0 or 1."),
            ],
            outputs=[gr.Video(label="output", autoplay=True)],
            allow_flagging="never",
            description="This is an unofficial demo for the paper 'FLATTEN: optical FLow-guided ATTENtion for consistent text-to-video editing'.",
            examples=[
                ["./data/puff.mp4", "A Tiger, high quality", "a cat with big eyes, deformed", 20, 16, 0],
                ["./data/background.mp4", "pointillism painting, detailed", "", 25, 16, 1],
                ["./data/trucks-race.mp4", "Wooden trucks drive on a racetrack.", "", 15, 16, 1],
            ],
            cache_examples=True,
        )
        demo.queue(max_size=10).launch()