app.py

from tuneavideo.pipelines.pipeline_tuneavideo import TuneAVideoPipeline
from tuneavideo.models.unet import UNet3DConditionModel
from tuneavideo.util import save_videos_grid
import torch
import gradio as gr


def tune_video_predict(
    pipe_id: str,
    prompt: str,
    video_length: int,
    height: int,
    width: int,
    num_inference_steps: int,
    guidance_scale: float,
):
    unet = UNet3DConditionModel.from_pretrained("Tune-A-Video-library/a-man-is-surfing", subfolder='unet', torch_dtype=torch.float16).to('cuda')
    pipe = TuneAVideoPipeline.from_pretrained(pipe_id, unet=unet, torch_dtype=torch.float16).to("cuda")
    video = pipe(prompt, video_length=video_length, height=height, width=width, num_inference_steps=num_inference_steps, guidance_scale=guidance_scale).videos
    output_path = save_videos_grid(video, save_path='output', path=f"{prompt}.gif")
    return output_path


demo_inputs = [
    gr.inputs.Dropdown(
        label="Model",
        choices=[
            "Tune-A-Video-library/a-man-is-surfing",
            "sd-dreambooth-library/mr-potato-head",
        ]
    ),
    gr.inputs.Textbox(
        label="Prompt",
        default='a flower blooming'

    ),
    gr.inputs.Slider(
        label="Video Length",
        minimum=1,
        maximum=50,
        default=8,
        step=1,
    ),
    gr.inputs.Slider(
        label="Height",
        minimum=128,
        maximum=1280,
        default=416,
        step=32,

    ),  
    gr.inputs.Slider(
        label="Width",
        minimum=128,
        maximum=1280,
        default=416,
        step=32,
    ),
    gr.inputs.Slider(
        label="Num Inference Steps",
        minimum=1,
        maximum=100,
        default=50,
        step=1,
    ),
    gr.inputs.Slider(
        label="Guidance Scale",
        minimum=0.0,
        maximum=100,
        default=7.5,
        step=0.5,
    )
]

demo_outputs = gr.outputs.Video(type="gif", label="Output")

examples = [
    ["Tune-A-Video-library/a-man-is-surfing", "a panda is surfing", 5, 416, 416, 50, 7.5],
    ["Tune-A-Video-library/a-man-is-surfing", "a flower blooming", 5, 416, 416, 50, 7.5],
    ["sd-dreambooth-library/mr-potato-head", "sks mr potato head, wearing a pink hat, is surfing.", 5, 416, 416, 50, 7.5],
    ["sd-dreambooth-library/mr-potato-head", "sks mr potato head is surfing in the forest.", 5, 416, 416, 50, 7.5],
]
    
description = "This generates video from an input text, using [one-shot tuning of diffusion models](https://arxiv.org/abs/2212.11565). To use it, simply input a text."

demo_app = gr.Interface(
    fn=tune_video_predict,
    inputs=demo_inputs,
    outputs=demo_outputs,
    examples=examples,
    cache_examples=False,
    title="Tune-A-Video",
    theme="huggingface",
    description=description
)

demo_app.launch(debug=True, enable_queue=True)