music-generator

ZeroX ACE Studio

A focused, local music-generation studio built on the official ACE-Step 1.5 Diffusers pipeline.

What it does

• Generates 48 kHz stereo music from a production prompt and structured lyrics.
• Supports Text-to-Music, Cover, Repaint, Extract, Lego, and Complete task inputs.
• Offers XL Turbo, XL SFT, and XL Base checkpoints.
• Saves every result as a downloadable 24-bit WAV.
• Includes six built-in lyric formulas and a one-click structure/syllable checker.
• Records the model, mode, seed, duration, and inference settings with each result.
• Uses VAE tiling and optional CPU offload to reduce GPU-memory pressure.

Honest hardware guide

The first launch downloads several gigabytes of model weights. Keep at least 20 GB of free disk space for experimentation with multiple XL models.

Windows quick start

1. Install 64-bit Python 3.12.
2. Put this folder somewhere with free disk space.
3. Double-click start_windows.bat.
4. Open http://127.0.0.1:7860 if it does not open automatically.

For the best NVIDIA performance, install the PyTorch build matching your installed GPU driver from the official PyTorch installer before starting the app.

Linux / macOS quick start

chmod +x start_unix.sh
./start_unix.sh

Manual start

python -m venv .venv
# Windows: .venv\Scripts\activate
# Linux/macOS: source .venv/bin/activate
python -m pip install --upgrade pip
pip install -r requirements.txt
python app.py

How to get better songs

Describe the complete sound, not only a genre. A strong prompt names the mood, tempo, instruments, vocal style, arrangement, and production character:

Cinematic deathcore fused with dark trap, Drop F guitars, 145 BPM, tense verses, huge half-time breakdown, industrial sub-bass, close male harsh vocals, wide modern master.

Structure lyrics with tags such as [intro], [verse], [pre-chorus], [chorus], [bridge], and [outro]. Reuse a seed when you want to recreate the same underlying generation.

Built-in lyrics formula

Choose a formula, click Load formula, and write inside the empty sections. Click Check lyrics before generating. The checker looks for a clear section order, a repeatable chorus, very long or short English lines, and the approximate 2,048-token ceiling. Its syllable count is a writing aid, not a perfect linguistic measurement.

ACE-Step's official guidance recommends roughly 6–10 syllables per line, concise combined tags such as [Chorus - anthemic], parentheses for background vocals, and blank lines between sections. Keep complex production details in the Music direction field so the model does not mistake them for words to sing.

Model behavior

• XL Turbo is guidance-distilled. The app fixes it to 8 steps and disables CFG by using guidance 1.0.
• XL SFT is a higher-step standard-generation model. Try 30–60 steps.
• XL Base supports the full advanced task set, including Extract, Lego, and Complete.

Rights and responsible release

The ACE-Step model is published under MIT terms and its authors describe the training corpus as licensed, royalty-free/public-domain, and synthetic. That does not guarantee that every generated result is unique or clear of third-party rights. Before a commercial release, check the output for recognizable melodies, lyrics, voices, trademarks, and samples. Only upload source audio you own or are authorized to transform.

Verified sources

• ACE-Step 1.5 official repository
• ACE-Step pipeline documentation
• Official XL Turbo Diffusers checkpoint
• ACE-Step 1.5 paper

This project is an independent interface and is not an official ACE-Step product.