The full ACE-Step engine on your own 4090 — simple on top, every dial underneath, and it teaches you each one.
checking engine…
Surprise me asks the model for a random style. Auto-format rewrites your style + lyrics into the caption format the model was trained on (and suggests bpm/key/duration) — usually a free quality win.
The model leans hard on specific, detailed descriptions (words like sub bass, clap backbeat, polished) — vague prompts sound vague. These are hand-tuned to sound good; tap one, tweak it, add your lyrics.
—
Cover art is generated after the song finishes and lands on the song in My songs a few seconds later.
Advanced — every dial, explained
Sound
Musical guardrails. Leave anything blank/0 and the model decides for itself — set it only when you know what you want.
Tempo in beats per minute. 90–110 = laid-back, 120–130 = dance, 140+ = drum&bass/hyper.
Locks the song's key. Minor = moodier, major = brighter. Blank lets the model pick what fits.
4/4 = almost all pop/rap. 3/4 = waltz feel. 6/8 = swung ballads.
Generation quality
How hard the model works per song. More steps = slower + usually cleaner, but ONLY on models built for it — the turbo models are trained to finish in 8.
How many refinement passes the model makes. Turbo models: 8 (more is wasted). SFT/merge models: ~50. Picking a model fills the right number in for you.
How strictly it follows your style prompt. Higher = more obedient but can sound strained; lower = freer. Turbo models ignore this entirely (they run without CFG).
Reproducibility
Same seed + same settings = the same song again. That's how you iterate: keep the seed, nudge one thing.
Every finished song shows its seed in My songs → ↺ reuse. Paste one here to rework that exact take.
Generate several variations in one go and keep the best. Each take costs the same GPU time again.
Output polish
Mastering touches applied to the finished file — safe to leave as-is.
Seconds of volume ramp at the start.
Seconds of ramp-down at the end — kills abrupt endings.
Evens the volume to a target dB so tracks sit level in a playlist. Streaming standard is -14.
Deeper experimental dials (DCW, FlowEdit, retake-variance, sampler/scheduler overrides) exist in the engine but aren't surfaced here yet — they're half-documented upstream, so I'm not going to pretend to teach them. They're on the roadmap once proven.
Everything you've generated, newest first — play, download, or ↺ reuse a song's exact settings.
Loading…
⬆ Upload a reference track (tap here) Give the model a real song to study: 🎯 use as reference steers the generation towards its vibe, and 📝 transcribe (local Whisper) pulls its lyrics out for you to remix.
Loading…
Open a song and shape it yourself — trim the top and tail, cut a bad bit, fade the ends, even out the volume, or stitch two takes together. It never touches your originals; you get a fresh file when you export. Works on generated songs and any audio file you open — even with the engine off.
—
0:00drag across the wave to select a section
Cut
Select a section first (drag across the wave). Trim keeps only what you selected; Delete removes it and closes the gap.
Fades
Smooths a start or ending so it doesn't jump in or cut off. Uses your selection, or the first / last 3 seconds if nothing's selected.
Volume
Slide to set a level, then Apply (to your selection, or the whole track). Normalise pushes it to a clean, consistent loudness automatically — the easy option.
100%
Speed
Nudges the whole track faster or slower. Like a record, this shifts the pitch too — small steps only, or it sounds silly.
Join
Stick a second song or take onto the end of this one — handy for gluing two takes into one track.