ACE-Step.

What to specify before generation

ACE-Step turns a description of the music and song lyrics into audio. In the interface, Caption describes the composition: its genre, instruments, mood and sense of tempo. The lyrics go in a separate Lyrics field. This lets you compare arrangements for the same lyrics or change the words while keeping the musical description. Duration, tempo, key and time signature can also be set with separate parameters.

There is an Instrumental switch for instrumental compositions. Before generating a song with vocals, choose the language and check how the words, endings and transitions between sections sound. The authors of the 1.5 branch report support for lyrics in more than 50 languages. This does not guarantee equal pronunciation or vocal quality across every language and genre.

The mode depends on the model you choose

In Custom mode, you can enter the description and lyrics yourself. An audio reference helps guide the style and timbre. Remix reworks a source recording, while Repaint lets you select a section to regenerate. After an edit, listen closely to the join with the retained audio: replacing a section can create an obvious transition or alter the sound.

The family has several model variants with different capabilities. The guide lists Extract, Lego and Complete for Base: extracting a part, adding an instrument and completing an incomplete arrangement. Turbo and SFT do not offer these modes. Before preparing source material, check the desired mode against the capability table for the specific variant. Versions 1.5 and XL belong to the same family but have different hardware requirements.

Example: a short background track

Suppose you need 30 seconds of calm background music for your own home repair video. You could describe soft keyboard tones, sparse percussion and steady dynamics, then enable Instrumental. Set the length separately with Audio Duration. Start with a short sketch and listen to it alongside the spoken track before deciding whether the chosen instruments work together.

When comparing variants, keep the description and duration unchanged and alter only the seed. Listen to the opening, transitions and ending. Check that the music does not compete with the speech and contains no unwanted vocals. Save the audio and the settings of a useful version for further work. The interface offers MP3, FLAC and WAV, among other formats. A file format alone does not confirm the quality of a composition.

Installation and use of the result

The code, weights and instructions are available for self-hosting. You can use a browser interface, a Python library or your own REST API server. Before installation, choose a compatible environment and model variant, taking memory, dependencies and device limitations into account. Your own API requires a running server and computing resources. Its availability does not mean the developers provide a ready-to-use cloud endpoint for every project.

The code and weights of the current 1.5 branch are published under MIT. Hardware, hosting and maintenance of your environment remain separate costs. The model license does not replace the rights needed for an uploaded recording, lyrics or other source material. Before distributing a result, check the material used and the recording itself. Research comparisons and speed claims on the project page should not be treated as results for your own computer without checking them.

Suitable tasks

• Create short instrumental sketches for your own videos
• Explore a musical idea with your own lyrics
• Create variants of your own recordings in a supported mode