Blog

September 15, 2026

Stems: Take Any Mix Apart

The Layers Teamextensionsstemsmixing

You have a rough mix and no session files. A reference track whose bassline you want to study. A demo where the vocal is great and everything else needs redoing. Every one of those starts the same way: you need the mix taken apart.

Stems is a Layers extension that does it in one command. Select a clip, run /stems:separate-selection from the chat, and six new clips appear right underneath it: drums, bass, vocals, other, guitar and piano.

S

Stems

Extension for Layers · Apple silicon Macs

Get the extension →

See a run, start to finish

This is the chat as it looks in the app, with the clock compressed. Click any stage to jump to it.

New chat
/stems:separate-selection

Started Stems: Separate Selection.

Stems

Downloading the separation model (336 MB), once.

Stems

Rendering Rough Mix v3.wav…Loading the model…

Stems

Split Rough Mix v3.wav into six stems: drums, bass, vocals, other, guitar and piano.

Rough Mix v3.wav
Ask about the selected clips…
Claude Opus 5 ⌄

Six real clips, one ⌘Z

The stems aren't a preview or a separate window — they're ordinary audio clips on your canvas, grouped under the source and in playing order. Move them, mute them, put effects on them, drag them into another idea.

The source clip is muted, not deleted, so you can unmute it any time to compare against the original. And the whole result — six clips, the group and the mute — goes in as a single edit. One ⌘Z and you're exactly where you started.

Separated dry, on purpose

Stems separates the clip as recorded, before its effect chain. If it separated the processed signal, your reverb would be baked into every stem — and playing those stems back through the same chain would apply it a second time. Dry in, dry out: your effects stay yours to reapply.

On your machine, not in the cloud

Separation runs on BS-RoFormer, a six-stem model that scores about 3 dB ahead of HTDemucs on every stem in the MVSEP multisong benchmark. It runs locally, on your Mac's CPU. Your audio is never uploaded.

The first run downloads the model — 336 MB, once — through Layers itself, checked against a SHA-256 that ships in the signed extension catalogue. Every run after that works offline.

What to expect

  • Roughly real time on Apple silicon: a three-minute song takes about three minutes. Progress shows live in the chat you ran it from.
  • One clip per run. With several selected, it names the one it's working on and leaves the rest alone.
  • Apple silicon Macs only for now — the model runtime has no Intel build.

Things to try

  • Pull the vocal off an old demo and build a new arrangement around it.
  • Mute the bass on a reference track and play along to learn the part.
  • Rebalance a bounced rough mix when the session files are long gone.
  • Lift a drum break out of a jam and chop it on the canvas.

Get it

In Layers, open the Extensions window, find Stems under Browse, and install it. It checks the signed catalogue and updates itself when a new version ships.

S

Stems

Extension for Layers · Apple silicon Macs

Get the extension →