app.py

import gradio as gr
import torch
import os
import uuid

from diffusers import AnimateDiffPipeline, MotionAdapter, EulerDiscreteScheduler
from diffusers.utils import export_to_video
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
from PIL import Image

SECRET_TOKEN = os.getenv('SECRET_TOKEN', 'default_secret')

# Constants
bases = {
    "ToonYou": "frankjoshua/toonyou_beta6",
    "epiCRealism": "emilianJR/epiCRealism"
}
step_loaded = None
base_loaded = "epiCRealism"
motion_loaded = None

# Ensure model and scheduler are initialized in GPU-enabled function
if not torch.cuda.is_available():
    raise NotImplementedError("No GPU detected!")

device = "cuda"
dtype = torch.float16
pipe = AnimateDiffPipeline.from_pretrained(bases[base_loaded], torch_dtype=dtype).to(device)
pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing", beta_schedule="linear")

def generate_image(secret_token, prompt, base, motion, step):
    if secret_token != SECRET_TOKEN:
        raise gr.Error(
            f'Invalid secret token. Please fork the original space if you want to use it for yourself.')


    global step_loaded
    global base_loaded
    global motion_loaded
    # print(prompt, base, step)

    if step_loaded != step:
        repo = "ByteDance/AnimateDiff-Lightning"
        ckpt = f"animatediff_lightning_{step}step_diffusers.safetensors"
        pipe.unet.load_state_dict(load_file(hf_hub_download(repo, ckpt), device=device), strict=False)
        step_loaded = step

    if base_loaded != base:
        pipe.unet.load_state_dict(torch.load(hf_hub_download(bases[base], "unet/diffusion_pytorch_model.bin"), map_location=device), strict=False)
        base_loaded = base

    if motion_loaded != motion:
        pipe.unload_lora_weights()
        if motion != "":
            pipe.load_lora_weights(motion, adapter_name="motion")
            pipe.set_adapters(["motion"], [0.7])
        motion_loaded = motion

    progress((0, step))
    def progress_callback(i, t, z):
        progress((i+1, step))

    output = pipe(
        prompt=prompt,

        # this corresponds roughly to 16:9
        # which is the aspect ratio video used by AiTube
        width=912, # 1024,
        height=512, # 576,
        
        guidance_scale=1.0,
        num_inference_steps=step,
        callback=progress_callback,
        callback_steps=1
    )

    name = str(uuid.uuid4()).replace("-", "")
    path = f"/tmp/{name}.mp4"
    
    # I think we are looking time here too, converting to mp4 is too slow, we should return
    # the frames unencoded to the frontend renderer
    export_to_video(output.frames[0], path, fps=10)

    # Read the content of the video file and encode it to base64
    with open(path, "rb") as video_file:
        video_base64 = base64.b64encode(video_file.read()).decode('utf-8')

    # Prepend the appropriate data URI header with MIME type
    video_data_uri = 'data:video/mp4;base64,' + video_base64
    
    # clean-up (otherwise there is always a risk of "ghosting", eg. someone seeing the previous generated video",
    # of one of the steps go wrong)
    os.remove(path)

    return video_data_uri


# Gradio Interface
with gr.Blocks() as demo:
    gr.HTML("""
        <div style="z-index: 100; position: fixed; top: 0px; right: 0px; left: 0px; bottom: 0px; width: 100%; height: 100%; background: white; display: flex; align-items: center; justify-content: center; color: black;">
        <div style="text-align: center; color: black;">
        <p style="color: black;">This space is a REST API to programmatically generate MP4 videos for AiTube, the next generation video platform.</p>
        <p style="color: black;">Interested in using it? Look no further than the <a href="https://huggingface.co/spaces/ByteDance/AnimateDiff-Lightning" target="_blank">original space</a>!</p>
        </div>
        </div>""")
    

    secret_token = gr.Text(label='Secret Token', max_lines=1)
    
    with gr.Group():
        with gr.Row():
            prompt = gr.Textbox(
                label='Prompt'
            )
        with gr.Row():
            select_base = gr.Dropdown(
                label='Base model',
                choices=[
                    "ToonYou", 
                    "epiCRealism",
                ],
                value=base_loaded,
                interactive=True
            )
            select_motion = gr.Dropdown(
                label='Motion',
                choices=[
                    ("Default", ""),
                    ("Zoom in", "guoyww/animatediff-motion-lora-zoom-in"),
                    ("Zoom out", "guoyww/animatediff-motion-lora-zoom-out"),
                    ("Tilt up", "guoyww/animatediff-motion-lora-tilt-up"),
                    ("Tilt down", "guoyww/animatediff-motion-lora-tilt-down"),
                    ("Pan left", "guoyww/animatediff-motion-lora-pan-left"),
                    ("Pan right", "guoyww/animatediff-motion-lora-pan-right"),
                    ("Roll left", "guoyww/animatediff-motion-lora-rolling-anticlockwise"),
                    ("Roll right", "guoyww/animatediff-motion-lora-rolling-clockwise"),
                ],
                value="",
                interactive=True
            )
            select_step = gr.Dropdown(
                label='Inference steps',
                choices=[
                    ('1-Step', 1), 
                    ('2-Step', 2),
                    ('4-Step', 4),
                    ('8-Step', 8)],
                value=4,
                interactive=True
            )
            submit = gr.Button()

    output_video_base64 = gr.Text()

    submit.click(
        fn=generate_image,
        inputs=[secret_token, prompt, select_base, select_motion, select_step],
        outputs=output_video_base64,
    )

app.queue(max_size=12).launch(show_api=True)