Austin French
← Playground

Music generation

Full songs with vocals from lyrics and style tags, using ACE-Step.

This tool ran on a GPU at my place, and it's been offline since I moved to New York. Below is how it worked. The engineering behind all of them is in austn.net's GPU tools.

model
ACE-Step on ComfyUI
An older version of the music tool, when the slider went to 8 minutes. Later versions capped tracks at four minutes.
An older version of the music tool, when the slider went to 8 minutes. Later versions capped tracks at four minutes.

The tool that surprised people most. You write lyrics with structure tags like [verse], [chorus] and [bridge], add style tags ("dark, death metal, electric guitar, 140 BPM, A minor"), pick a length, and get back a whole song with vocals.

How it worked

  • Presets (instrumental, full song, rap, ambient) fill in sensible tags and settings, and anything you set yourself wins.
  • Duration, inference steps and the guidance scales are clamped on the server before anything reaches the GPU. Length tops out at four minutes.
  • A song can take minutes to render, so the job's hold on the GPU lock is sized to the longest song it allows. Otherwise the lock could expire mid-song and let a second job onto the GPU.
  • The finished audio could be downloaded as a file.