ACE-Step: three minutes of music in 27 seconds, on your own card
ACE-Step is an open-source model that writes music from style tags and, if you want, lyrics. Its strength is speed: on our card, three minutes of music in 27 seconds. Fast enough to generate twenty candidates, listen, and keep three. It also reworks a passage without touching the rest.
What you get. Royalty-free music made at home: ambience tracks for focus, soundtracks for videos, sketches for songs. No subscription, no upload, and a bank of finished tracks that plays later with no model at all.
1. What it does
- Generate: a full track from tags (
soft piano and warm pads, meditative, no drums, no vocals). - Repaint: redo a time range, the rest stays identical.
- Retake: a variation of the whole, same idea, different take.
Run it in its own Python environment: its PyTorch and CUDA versions rarely match those of your other tools. We call it through a small bridge, the single place that adapts to its interface when it changes (see fh-seq).
2. Three uses here
- An ambience bank. Twenty instrumental tracks per style in one call, listened to, sorted, the best kept. Purrd then plays them while an agent works, with no model running.
- Video soundtracks. In our automatic video chain, the music for a three-minute film takes 27 seconds, next to minutes per video shot.
- Song sketches, reworked passage by passage.
3. Generate before, play without AI
The structuring choice: the model only runs ahead of time. The program that plays the music reads files, at random, with zero graphics memory and zero dependency on a model. The AI makes the bank; nothing in the playback path needs it.
Traps
Size the storage: 32-bit stereo WAV at 48 kHz, three minutes weigh 69 MB. Converted to Opus at 128 kbit/s: about 3 MB, no audible loss for ambience.
Know which card it renders on. We believed it shared nothing with our language models; it rendered on the same card, and a model started during a render stayed on the processor for an hour after.
Tags do the work. For an "end of task" track, words like resolving to silence, decaying piano notes gave what we wanted; vague tags give vague music.
Sources
Found this useful? Tell us what to measure next.