The gist

ACE-Step 1.5 is an open-source music foundation model from ACE Studio and StepFun. It provides commercial-grade music generation that runs efficiently on consumer hardware. Its hybrid architecture uses a Language Model to plan comprehensive song blueprints from simple user queries, which a Diffusion Transformer then uses to synthesize the final audio, complete with lyrics and metadata.

What it does

Generate full songs, from 10 seconds to 10 minutes long, from text descriptions.
Create cover versions of songs using reference audio to guide the style.
Edit audio by repainting sections, separating tracks into stems, or generating accompaniment.
Train a custom LoRA model on a small set of songs to capture a unique musical style.
Control specific musical attributes such as BPM, key, scale, and time signature.
Generate lyrics with timestamps in over 50 languages.

How it works

Users install ACE-Step 1.5 by cloning its GitHub repository and running local setup scripts. It can be operated via a Gradio web UI, a REST API, or a command-line interface. Users provide text prompts, and optionally reference audio, to guide generation, and the model outputs a complete audio file. As an open-source project under the MIT license, it runs on a user's own GPU and is free to use.

Best for

ACE-Step 1.5 is best for developers, music producers, and researchers who need a powerful, locally-hosted music generation model. It is the right choice for users requiring deep customization, such as training their own LoRA models, or programmatic access via its API.

Watch out for

While it can run on lower-end GPUs, achieving the highest audio quality with the XL models requires significant VRAM (12-20GB+). Users must carefully select the correct model components and configuration for their specific hardware to balance performance and quality.