ACE-Step 1.5
A local, open-source music generator that can turn a text prompt and lyrics into a full song on your own computer.
- Type AI generation
- Difficulty Medium
- Hardware Python 3.11-3.12. A CUDA GPU is recommended; Apple MPS, AMD ROCm, Intel XPU, and CPU are also supported. Official install docs: at least 4 GB of VRAM for DiT-only mode, at least 6 GB for LM+DiT. The later XL (4B DiT) models need at least 12 GB of VRAM with offload; 20 GB is recommended. About 10 GB of disk for the core models. CPU works for inference only and is much slower.
- License MIT
What it does
ACE-Step 1.5 is an open-source music foundation model from ACE Studio and StepFun. You describe a style, add lyrics if you want them, and it generates a complete piece of music locally. The project says duration can run from 10 seconds up to 10 minutes, with tools for covers, selective repaint edits, stem separation, multi-track layering, and building accompaniment under a vocal.
It is not a DAW. You get audio out, then you edit in whatever you already use. There is a local Gradio UI and a REST API option. Models download on first run.
The project also points people who do not want a local install to acemusic.ai, which it describes as a free hosted option. The README says the only official project website is their GitHub Pages site and warns that look-alike domains exist. Stick to the official repo and the links it names.
Why it matters for hip-hop
A local generator is useful when you want a hook sketch, a reference beat, or a topline idea without uploading your lyrics to a subscription app. You can prompt for rap or R&B cadences, try several versions, and keep the ones that feel close. The project also says you can train a light LoRA from a few of your own songs, and that lyrics can be written in many languages. Treat the output as a sketch until you have replaced or replayed what you do not own musically.
Getting it running
The official quick start is: install uv, clone the repo, run uv sync, then launch the app. That is straightforward if you are comfortable in a terminal, which is why this is Medium rather than Easy. The first model download takes time and disk. If you only have a modest GPU, start with DiT-only or turbo settings rather than the XL models.
License notes
The repository license is the MIT License. The official Hugging Face model card for ACE-Step 1.5 also marks the model as MIT and states that generated music can be used for commercial purposes. The same card says the training data was built from licensed tracks, public-domain or royalty-free music, and synthetic audio. That is what the project itself says, not a legal opinion. Read the current license before you release anything. Using a copyrighted song as a reference, or asking the model to imitate a specific artist too closely, are separate copyright questions. The project's README asks users to check originality, disclose AI involvement where it matters, and respect the law.