---
title: "ACE-Step v1.5 XL Turbo"
sdk: gradio
app_file: app.py
python_version: "3.10"
short_description: "Text-to-music with ACE-Step v1.5 XL Turbo"
---

# ACE-Step v1.5 XL Turbo Diffusers

A private Hugging Face Space that generates music from text prompts using [ACE-Step v1.5 XL Turbo](https://huggingface.co/ACE-Step/acestep-v15-xl-turbo-diffusers).

## What it does

- **Text-to-music generation**: Provide a prompt describing the music you want.
- **Optional lyrics**: Supply lyrics to guide the musical generation.
- **Controllable duration**: Generate from 5 to 120 seconds of audio.
- **Seed control**: Reproducible generation with a user-provided seed.

## How to try

1. Open the **Generate** tab.
2. Type a prompt, e.g.:
   - *"An upbeat synthwave track with driving drums and a catchy lead"*
3. Optionally add lyrics.
4. Set duration and seed.
5. Click **Generate** and wait for the output audio to appear.

## Example inputs

- **Prompt**: `An upbeat synthwave track with driving drums and a catchy lead`
- **Lyrics** (optional): `[Verse]\nNeon lights are calling me\n[Chorus]\nRide the wave tonight`
- **Duration**: `30`
- **Seed**: `42`

## Technical notes

- The model is a 5B-parameter flow-matching DiT in the `diffusers` format.
- It loads at module level with `bfloat16` precision on CUDA.
- VAE tiling is enabled to keep memory bounded during decode.
- Default inference uses 8 denoising steps (`guidance_scale > 1.0` is ignored because this is a guidance-distilled turbo checkpoint).
- The app is designed for **ZeroGPU** (`zero-a10g`) with a fixed GPU fallback; the inference function is wrapped with `@spaces.GPU`.

## Dependencies

- `diffusers` (from source until the next release includes `AceStepPipeline`)
- `gradio`, `spaces`, `torchaudio`, `soundfile`, `accelerate`
- `torch` is provided by the platform layer and must not be pinned.

## Limitations

- First-time generation after a cold start may take longer due to model download and weight loading.
- If the Space cannot load the pipeline (e.g. missing symbol or OOM), the **Health** endpoint reports `pipeline_ready=false` and the concrete error.