ACE-Step v1.5

ACE-Step v1.5

Generates full songs from simple descriptions and lyrics, and edits or covers existing audio, with the ACE-Step 1.5 music model. It suits musicians and creators who want local, fast AI music tools.

Made by ACE-Step on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: This Space is a demo of ACE-Step 1.5, an open-source music generation model built to run on consumer hardware. A language model plans each song from a short request, and a diffusion transformer then makes the audio. It can also cover, repaint and edit audio. The demo runs on zero-a10g hardware at Hugging Face.

Who it suits: It suits musicians, producers and creators who want to draft songs quickly and edit audio, and developers who want a music model they can run on their own machine.

Industry | Who uses it | How they use it | Example
Music production | Producers and songwriters | Draft songs from a short description | Ask for a full song and let the model plan the structure
Content creation | Video and podcast makers | Make tracks for a project | Add an accompaniment to a recorded vocal
Sound editing | Audio editors | Change parts of an existing track | Repaint one section of a generated song
AI development | Developers | Run and fine-tune a music model locally | Train a LoRA from a few songs

You can skip it if you want a hosted service with a polished studio interface, since this is a model demo.

Where to use it: You can use it in these places:

• A browser: open the Space's page on Hugging Face and nothing needs installing.
• Windows, macOS or Linux: run your own copy with Python 3.11.
• A local Gradio web interface, started with the project's own commands.

Use case | Who | What they get back
Song generation | Musicians | A song from a description, from 10 seconds up to 10 minutes
Cover and edit | Producers | A cover or a repainted section of existing audio
Accompaniment | Vocalists | Backing music generated for a vocal track

How to set it up: You need these first for a local copy:

• Python 3.11 and pip
• A CUDA GPU is recommended, and the README says CPU or MPS also work but slower
• The system packages listed in packages.txt, which Hugging Face installs with apt-get

The repository also holds a Dockerfile and an .env.example file for settings. The online page needs none of this. For your own copy of the Space, follow the steps below.

1. Install the Hugging Face tools.
2. Download the Space's files into a folder.
3. Move into that folder.
4. Install Gradio at the version the Space uses, then the other packages.
5. Start the app.

pip install -U huggingface_hub
hf download ACE-Step/Ace-Step-v1.5 --repo-type space --local-dir Ace-Step-v1.5
cd Ace-Step-v1.5
pip install gradio==6.2.0
pip install -r requirements.txt
python app.py

The project's own route installs the uv package manager and clones its repository. Then it runs these two commands inside the project folder.

uv sync
uv run acestep

How to use it: Open the Space in a browser, or start your own copy. Describe the song you want, add lyrics if you have them and run the generator. You can also give reference audio to steer the style.

Tool | What it does
Reference audio input | Uses an audio clip to steer the style of the result
Cover generation | Makes a cover from existing audio
Repaint and edit | Regenerates a chosen part of the audio
Track separation | Splits audio into separate stems
Multi-track generation | Adds new layers to a track
Vocal2BGM | Makes accompaniment for a vocal track
Metadata control | Sets duration, BPM, key, scale and time signature
Simple mode | Builds a full song from a plain description
Audio understanding | Reads BPM, key, time signature and a caption from audio
LRC generation | Makes lyric timestamps for generated music
LoRA training | Annotates and trains in Gradio, in about an hour for 8 songs on a 12 GB card
Quality scoring | Rates the quality of generated audio

Example: You type a plain description of a song you want, and choose a duration. The model plans the lyrics and structure, then returns a generated song that follows the description.

For newcomers and experts: If you are new to this, try Simple mode with one sentence about the song you want. Listen to the result, then change the sentence and run it again.

If you know the field, a language model acts as a planner that writes song blueprints, lyrics and captions through chain-of-thought, and a diffusion transformer renders the audio. The authors say alignment comes from the model's own internal signals instead of an outside reward model. They report generation in 0.5 to 10 seconds on an A100 and up to 8 songs per batch.

You describe the song
The language model plans it
The diffusion transformer renders audio
You edit, cover or repaint

Good to know: The README says the model is tuned for local use with under 4 GB of VRAM. The audio can run from 10 seconds to 10 minutes, and the authors list support for more than 50 languages.

• License: mit
• Cost to run: none named
• Data it reaches: your prompts, lyrics and any reference audio you give, plus model files downloaded from Hugging Face
• Not for: GPU-free use at full speed, since CPU and MPS run slower

On Hugging Face, 13 November 2026

• Likes: 596
• Built with: gradio
• License: mit

Links

• Space: https://huggingface.co/spaces/ACE-Step/Ace-Step-v1.5
• Files: https://huggingface.co/spaces/ACE-Step/Ace-Step-v1.5/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.