app.py

import spaces
import torch
from diffusers import HunyuanVideoPipeline, HunyuanVideoTransformer3DModel
# from para_attn.first_block_cache.diffusers_adapters import apply_cache_on_pipe
from PIL import Image
import numpy as np
import gradio as gr
import os


# device = "cuda" if torch.cuda.is_available() else "cpu"

# model_id = "hunyuanvideo-community/HunyuanVideo"

model_id = "FastVideo/FastHunyuan-diffusers"

transformer = HunyuanVideoTransformer3DModel.from_pretrained(
    model_id, subfolder="transformer", torch_dtype=torch.bfloat16
    )


pipe = HunyuanVideoPipeline.from_pretrained(model_id, transformer=transformer, torch_dtype=torch.float16)
pipe.vae.enable_tiling()
# pipe.load_lora_weights("")
# pipe.to("cuda")

@spaces.GPU()
def generate(prompt, width=832, height=832, num_inference_steps=30, lora_id=None, progress=gr.Progress(track_tqdm=True)):
    if lora_id and lora_id.strip() != "":
        pipe.unload_lora_weights()
        pipe.load_lora_weights(lora_id.strip())
    # apply_cache_on_pipe(
    #     pipe,
    #     # residual_diff_threshold=0.2,
    # )
    pipe.to("cuda")
    torch.cuda.empty_cache()
    try:
        output = pipe(
            prompt=prompt,
            # negative_prompt=negative_prompt,
            height=height,
            width=width,
            num_frames=1,
            num_inference_steps=num_inference_steps,
            # guidance_scale=5.0,
        ).frames[0][0]
        # image = (output * 255).astype(np.uint8)
        # return Image.fromarray(image)
        return output
    finally:
        # Always clear memory, even if an error occurs
        if lora_id and lora_id.strip() != "":
            pipe.unload_lora_weights()
        torch.cuda.empty_cache()


iface = gr.Interface(
    fn=generate,
    inputs=[
        gr.Textbox(label="Input prompt"),
    ],
    additional_inputs = [
        # gr.Textbox(label="Negative prompt", value = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"),
        gr.Slider(label="Width", minimum=192, maximum=1280, step=16, value=832),
        gr.Slider(label="Height", minimum=192, maximum=1280, step=16, value=832),
        gr.Slider(minimum=1, maximum=80, step=1, label="Inference Steps", value=10),
        gr.Textbox(label="LoRA ID"),
    ],
    outputs=gr.Image(label="output"),
)


iface.launch(share=True, debug=True)