app.py

import spaces

import os
import random

import gradio as gr
import numpy as np
import torch
from diffusers import StableAudioPipeline


MODEL_ID = "alfredplpl/SAO-Re-2"
token=os.environ["HF_TOKEN"]

device = "cuda"
dtype = torch.bfloat16

pipe = StableAudioPipeline.from_pretrained(
    MODEL_ID,
    torch_dtype=dtype,
    token=token
)
pipe = pipe.to(device)

if device == "cuda":
    pipe.enable_model_cpu_offload()

@spaces.GPU
def generate_audio(
    prompt: str,
    negative_prompt: str,
    duration: float,
    steps: int,
    guidance_scale: float,
    seed: int,
):
    if not prompt.strip():
        raise gr.Error("Promptを入力してください。")

    if seed < 0:
        seed = random.randint(0, 2**32 - 1)

    generator = torch.Generator(device=device).manual_seed(seed)

    with torch.inference_mode():
        audio = pipe(
            prompt=prompt,
            negative_prompt=negative_prompt or None,
            num_inference_steps=steps,
            guidance_scale=guidance_scale,
            audio_end_in_s=duration,
            num_waveforms_per_prompt=1,
            generator=generator,
        ).audios[0]

    # diffusers StableAudio: [channels, samples] torch tensor の想定
    audio = audio.T.float().cpu().numpy()

    # Gradio向けに軽く安全化
    audio = np.nan_to_num(audio)
    peak = np.max(np.abs(audio))
    if peak > 1.0:
        audio = audio / peak

    sample_rate = pipe.vae.sampling_rate

    return (sample_rate, audio), seed


with gr.Blocks(title="Text-to-Audio") as demo:
    gr.Markdown(
        """
# Text-to-Audio Demo

Text-to-Audio Demo of Stable Audio Open ReImplementation. You may use this output for commercial purposes.
"""
    )

    with gr.Row():
        with gr.Column():
            prompt = gr.Textbox(
                label="Prompt",
                value="A woman is speaking.",
                lines=2,
            )
            negative_prompt = gr.Textbox(
                label="Negative Prompt",
                value="low quality, quality, music",
                lines=2,
            )

            duration = gr.Slider(
                minimum=1.0,
                maximum=10.0,
                value=5.0,
                step=0.5,
                label="Duration seconds",
            )
            steps = gr.Slider(
                minimum=10,
                maximum=300,
                value=200,
                step=1,
                label="Inference steps",
            )
            guidance_scale = gr.Slider(
                minimum=1.0,
                maximum=15.0,
                value=7.0,
                step=0.5,
                label="Guidance scale",
            )
            seed = gr.Number(
                label="Seed -1 means random",
                value=-1,
                precision=0,
            )

            button = gr.Button("Generate", variant="primary")

        with gr.Column():
            output_audio = gr.Audio(
                label="Generated Audio",
                type="numpy",
            )
            used_seed = gr.Number(
                label="Used Seed",
                precision=0,
            )

    button.click(
        fn=generate_audio,
        inputs=[
            prompt,
            negative_prompt,
            duration,
            steps,
            guidance_scale,
            seed,
        ],
        outputs=[
            output_audio,
            used_seed,
        ],
    )

    gr.Examples(
        examples=[
             [
                "Playing a guitar.",
                "low quality, noisy",
                5.0,
                200,
                7.0,
                -1,
            ],
            [
                "Playing a piano.",
                "low quality, noisy",
                5.0,
                200,
                7.0,
                -1,
            ],
            [
                "wind",
                "",
                5.0,
                200,
                7.0,
                -1,
            ],
        ],
        inputs=[
            prompt,
            negative_prompt,
            duration,
            steps,
            guidance_scale,
            seed,
        ],
    )


if __name__ == "__main__":
    demo.queue().launch()