app.py

import spaces
import gradio as gr
import argparse
import sys
import time
import os
import random
from skyreelsinfer.offload import OffloadConfig
from skyreelsinfer import TaskType
from skyreelsinfer.skyreels_video_infer import SkyReelsVideoSingleGpuInfer
from diffusers.utils import export_to_video
from diffusers.utils import load_image

#predictor = None
#task_type = None

@spaces.GPU(duration=120)
#def init_predictor():
#    global predictor
predictor = SkyReelsVideoSingleGpuInfer(
        task_type= TaskType.I2V,
        model_id="Skywork/SkyReels-V1-Hunyuan-I2V",
        quant_model=True,
        is_offload=True,
        offload_config=OffloadConfig(
            high_cpu_memory=True,
            parameters_level=True,
            compiler_transformer=False,
        )
)
    
@spaces.GPU(duration=80)
def generate_video(prompt, seed, image=None):
    print(f"image:{type(image)}")
    if seed == -1:
        random.seed(time.time())
        seed = int(random.randrange(4294967294))
    kwargs = {
        "prompt": prompt,
        "height": 512,
        "width": 512,
        "num_frames": 97,
        "num_inference_steps": 30,
        "seed": seed,
        "guidance_scale": 6.0,
        "embedded_guidance_scale": 1.0,
        "negative_prompt": "Aerial view, aerial view, overexposed, low quality, deformation, a poor composition, bad hands, bad teeth, bad eyes, bad limbs, distortion",
        "cfg_for": False,
    }
    assert image is not None, "please input image"
    kwargs["image"] = load_image(image=image)
    #global predictor
    output = predictor.inference(kwargs)
    save_dir = f"./result/{task_type}"
    os.makedirs(save_dir, exist_ok=True)
    video_out_file = f"{save_dir}/{prompt[:100].replace('/','')}_{seed}.mp4"
    print(f"generate video, local path: {video_out_file}")
    export_to_video(output, video_out_file, fps=24)
    return video_out_file, kwargs

def create_gradio_interface():
        with gr.Blocks() as demo:
            with gr.Row():
                image = gr.Image(label="Upload Image", type="filepath")
                prompt = gr.Textbox(label="Input Prompt")
                seed = gr.Number(label="Random Seed", value=-1)
            submit_button = gr.Button("Generate Video")
            output_video = gr.Video(label="Generated Video")
            output_params = gr.Textbox(label="Output Parameters")
            submit_button.click(
                fn=generate_video,
                inputs=[prompt, seed, image],
                outputs=[output_video, output_params],
            )
        return demo

if __name__ == "__main__":
    #import multiprocessing
    #multiprocessing.freeze_support()
    #init_predictor()
    demo = create_gradio_interface()
    demo.launch()