File size: 5,028 Bytes
55e719f
67c95e3
086bafc
67c95e3
ecd0db9
 
 
6a96530
ecd0db9
7265fcb
0edac6b
 
 
 
 
 
 
6ed4aaa
 
0edac6b
 
 
 
67c95e3
8db257f
6ed4aaa
67c95e3
 
0396c9b
6ed4aaa
67c95e3
6ed4aaa
67c95e3
 
 
 
6ed4aaa
 
67c95e3
ecd0db9
49f3202
6ed4aaa
 
67c95e3
 
 
ecd0db9
 
67c95e3
 
 
ecd0db9
67c95e3
ecd0db9
67c95e3
ecd0db9
 
 
 
67c95e3
 
 
 
ecd0db9
 
 
f8a22eb
8db257f
ecd0db9
 
f8a22eb
ecd0db9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f8a22eb
ecd0db9
0edac6b
ecd0db9
 
 
 
 
 
 
 
 
 
 
 
 
 
f8a22eb
 
ecd0db9
8db257f
97ab6fc
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
import spaces
import torch
from diffusers import HunyuanVideoPipeline, HunyuanVideoTransformer3DModel

import os
import time
from datetime import datetime
import gradio as gr

@spaces.GPU
def generate_video(
    prompt,
    resolution,
    video_length,
    seed,
    num_inference_steps,
    guidance_scale,
    flow_shift, # TODO: change to flow_shift 
    embedded_guidance_scale # TODO: change to embedded_guidance_scale
):
    seed = None if seed == -1 else seed
    width, height = resolution.split("x")
    width, height = int(width), int(height)

    model = "hunyuanvideo-community/HunyuanVideo"
    transformer = HunyuanVideoTransformer3DModel.from_pretrained(
        model,
        subfolder="transformer",
        device_map="balanced",
        torch_dtype=torch.float16, # TODO: change to bfloat16
    )
    print(f"transformer device: {transformer.device}")

    # Cargar el pipeline
    pipeline = HunyuanVideoPipeline.from_pretrained(
        model,
        transformer=transformer,
        torch_dtype=torch.float16, # TODO: change to bfloat16
        device_map="balanced",
    )
    print(f"pipeline device: {pipeline.device}")
    # TODO: pipeline.vae.enable_tiling()
    # TODO: pipeline.to("cuda")
    
    # Generar el video usando el pipeline
    video = pipeline(
        prompt=prompt,
        height=height,
        width=width,
        num_frames=video_length,
        num_inference_steps=num_inference_steps,
        guidance_scale=guidance_scale,
    ).frames[0]
    
    # Guardar el video
    save_path = os.path.join(os.getcwd(), "gradio_outputs")
    os.makedirs(save_path, exist_ok=True)
    
    time_flag = datetime.fromtimestamp(time.time()).strftime("%Y-%m-%d-%H:%M:%S")
    video_path = f"{save_path}/{time_flag}_seed{seed}_{prompt[:100].replace('/','')}.mp4"
    
    from diffusers.utils import export_to_video
    export_to_video(video, video_path, fps=24)
    print(f'Sample saved to: {video_path}')
    
    return video_path

def create_demo():
    with gr.Blocks() as demo:
        gr.Markdown("# Hunyuan Video Generation")
        
        with gr.Row():
            with gr.Column():
                prompt = gr.Textbox(label="Prompt", value="A cat walks on the grass, realistic style.")
                with gr.Row():
                    resolution = gr.Dropdown(
                        choices=[
                            # 720p
                            ("1280x720 (16:9, 720p)", "1280x720"),
                            ("720x1280 (9:16, 720p)", "720x1280"), 
                            ("1104x832 (4:3, 720p)", "1104x832"),
                            ("832x1104 (3:4, 720p)", "832x1104"),
                            ("960x960 (1:1, 720p)", "960x960"),
                            # 540p
                            ("960x544 (16:9, 540p)", "960x544"),
                            ("544x960 (9:16, 540p)", "544x960"),
                            ("832x624 (4:3, 540p)", "832x624"), 
                            ("624x832 (3:4, 540p)", "624x832"),
                            ("720x720 (1:1, 540p)", "720x720"),
                        ],
                        value="1280x720",
                        label="Resolution"
                    )
                    video_length = gr.Dropdown(
                        label="Video Length",
                        choices=[
                            ("2s(65f)", 65),
                            ("5s(129f)", 129),
                        ],
                        value=129,
                    )
                num_inference_steps = gr.Slider(1, 100, value=50, step=1, label="Number of Inference Steps")
                show_advanced = gr.Checkbox(label="Show Advanced Options", value=False)
                with gr.Row(visible=False) as advanced_row:
                    with gr.Column():
                        seed = gr.Number(value=-1, label="Seed (-1 for random)")
                        guidance_scale = gr.Slider(1.0, 20.0, value=1.0, step=0.5, label="Guidance Scale")
                        flow_shift = gr.Slider(0.0, 10.0, value=7.0, step=0.1, label="Flow Shift") 
                        embedded_guidance_scale = gr.Slider(1.0, 20.0, value=6.0, step=0.5, label="Embedded Guidance Scale")
                show_advanced.change(fn=lambda x: gr.Row(visible=x), inputs=[show_advanced], outputs=[advanced_row])
                generate_btn = gr.Button("Generate")
            
            with gr.Column():
                output = gr.Video(label="Generated Video")
        
        generate_btn.click(
            fn=lambda *inputs: generate_video(*inputs),
            inputs=[
                prompt,
                resolution,
                video_length,
                seed,
                num_inference_steps,
                guidance_scale,
                flow_shift,
                embedded_guidance_scale
            ],
            outputs=output
        )
    
    return demo

if __name__ == "__main__":
    print("Starting Gradio server...")
    demo = create_demo()
    demo.launch()