demo_app.py

import spaces
import gc
import gradio as gr
import numpy as np
import os
from pathlib import Path
from diffusers import GGUFQuantizationConfig, HunyuanVideoPipeline, HunyuanVideoTransformer3DModel
from diffusers.utils import export_to_video
from huggingface_hub import snapshot_download
import torch

gc.collect()
torch.cuda.empty_cache()
torch.set_grad_enabled(False)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False

model_id = "hunyuanvideo-community/HunyuanVideo"
base_path = f"/home/user/app/{model_id}"
os.makedirs(base_path, exist_ok=True)
snapshot_download(repo_id=model_id, local_dir=base_path)
ckp_path = Path(base_path)

gguf_filename = "hunyuan-video-t2v-720p-Q4_0.gguf"
transformer_path = f"https://huggingface.co/city96/HunyuanVideo-gguf/blob/main/{gguf_filename}"
transformer = HunyuanVideoTransformer3DModel.from_single_file(
    transformer_path,
    quantization_config=GGUFQuantizationConfig(compute_dtype=torch.bfloat16),
    torch_dtype=torch.bfloat16,
)
transformer = transformer.to('cuda')

pipe = HunyuanVideoPipeline.from_pretrained(
    ckp_path,
    transformer=transformer,
    torch_dtype=torch.float16
)

if pipe.text_encoder:
    pipe.text_encoder = pipe.text_encoder.to('cuda')
    pipe.text_encoder.eval()

pipe.vae.enable_tiling()
pipe.vae.enable_slicing()
pipe.vae.eval()
pipe.vae = pipe.vae.to("cuda")
pipe = pipe.to("cuda")

pipe.load_lora_weights(
    "sergidev/IllustrationTTV",
    weight_name="hunyuan_flat_color_v2.safetensors",
    adapter_name="hyvid_lora_adapter"
)
pipe.set_adapters("hyvid_lora_adapter", 1.2)

gc.collect()
torch.cuda.empty_cache()

MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 1024

@spaces.GPU(duration=120)
def generate(
    prompt,
    height,
    width,
    num_frames,
    num_inference_steps,
    seed_value,
    fps,
    progress=gr.Progress(track_tqdm=True)
):
    with torch.cuda.device(0):
        if seed_value == -1:
            seed_value = torch.randint(0, MAX_SEED, (1,)).item()
        generator = torch.Generator('cuda').manual_seed(seed_value)

        with torch.amp.autocast_mode.autocast('cuda', dtype=torch.bfloat16), torch.inference_mode(), torch.no_grad():
            output = pipe(
                prompt=prompt,
                height=height,
                width=width,
                num_frames=num_frames,
                num_inference_steps=num_inference_steps,
                generator=generator,
            ).frames[0]

        output_path = "output.mp4"
        export_to_video(output, output_path, fps=fps)
        torch.cuda.empty_cache()
    gc.collect()
    return output_path

def apply_preset(preset_name, *current_values):
    if preset_name == "Higher Resolution":
        return [608, 448, 24, 29, 12]
    elif preset_name == "More Frames":
        return [512, 320, 42, 27, 14]
    return current_values

css = """
#col-container {
    margin: 0 auto;
    max-width: 850px;
}

.dark-theme {
    background-color: #1f1f1f;
    color: #ffffff;
}

.container {
    margin: 0 auto;
    padding: 20px;
    border-radius: 10px;
    background-color: #2d2d2d;
    box-shadow: 0 4px 6px rgba(0, 0, 0, 0.1);
}

.title {
    text-align: center;
    margin-bottom: 1em;
    color: #ffffff;
}

.description {
    text-align: center;
    margin-bottom: 2em;
    color: #cccccc;
    font-size: 0.95em;
    line-height: 1.5;
}

.prompt-container {
    background-color: #363636;
    padding: 15px;
    border-radius: 8px;
    margin-bottom: 1em;
    width: 100%;
}

.prompt-textbox {
    min-height: 80px !important;
}

.preset-buttons {
    display: flex;
    gap: 10px;
    justify-content: center;
    margin-bottom: 1em;
}

.support-text {
    text-align: center;
    margin-top: 1em;
    color: #cccccc;
    font-size: 0.9em;
}

a {
    color: #00a7e1;
    text-decoration: none;
}

a:hover {
    text-decoration: underline;
}
"""

with gr.Blocks(css=css, theme="dark") as demo:
    with gr.Column(elem_id="col-container"):
        gr.Markdown("# 🎬 Illustration TTV", elem_classes=["title"])
        gr.Markdown(
            """Transform your text descriptions into illustrative videos using HunyuanVideo for free!
            This space uses the 'hunyuan flat color v2' LORA by Motimalu to generate better 2d animated sequences. Prompt only handles 77 tokens.

            If you find this useful, please consider giving the space a ❤️ and supporting me on [Ko-Fi](https://ko-fi.com/sergidev)!""",
            elem_classes=["description"]
        )

        with gr.Column(elem_classes=["prompt-container"]):
            prompt = gr.Textbox(
                label="Prompt",
                placeholder="Enter your prompt here (Include the terms 'flat color, no lineart, blending' for 2d illustration)",
                show_label=False,
                elem_classes=["prompt-textbox"],
                lines=3
            )

        with gr.Row():
            run_button = gr.Button("🎨 Generate", variant="primary", size="lg")

        with gr.Row(elem_classes=["preset-buttons"]):
            preset_high_res = gr.Button("📺 Higher Resolution Preset")
            preset_more_frames = gr.Button("🎞️ More Frames Preset")

        with gr.Row():
            result = gr.Video(label="Generated Video")

        with gr.Accordion("⚙️ Advanced Settings", open=False):
            seed = gr.Slider(
                label="Seed (-1 for random)",
                minimum=-1,
                maximum=MAX_SEED,
                step=1,
                value=-1,
            )
            with gr.Row():
                height = gr.Slider(
                    label="Height",
                    minimum=256,
                    maximum=MAX_IMAGE_SIZE,
                    step=16,
                    value=608,
                )
                width = gr.Slider(
                    label="Width",
                    minimum=256,
                    maximum=MAX_IMAGE_SIZE,
                    step=16,
                    value=448,
                )
            with gr.Row():
                num_frames = gr.Slider(
                    label="Number of frames to generate",
                    minimum=1.0,
                    maximum=257.0,
                    step=1,
                    value=24,
                )
                num_inference_steps = gr.Slider(
                    label="Number of inference steps",
                    minimum=1,
                    maximum=50,
                    step=1,
                    value=29,
                )
            fps = gr.Slider(
                label="Frames per second",
                minimum=1,
                maximum=60,
                step=1,
                value=12,
            )

    # Event handling
    run_button.click(
        fn=generate,
        inputs=[prompt, height, width, num_frames, num_inference_steps, seed, fps],
        outputs=[result],
    )

    # Preset button handlers
    preset_high_res.click(
        fn=lambda: apply_preset("Higher Resolution"),
        outputs=[height, width, num_frames, num_inference_steps, fps]
    )

    preset_more_frames.click(
        fn=lambda: apply_preset("More Frames"),
        outputs=[height, width, num_frames, num_inference_steps, fps]
    )