Getting the best results

This isn't a full manual — every control in the app is labeled, and most have a tooltip. This page is about the decisions that matter more than any one slider: when to reach for what, and why.

Start here: drop a track, or load pre-split stems?

If you have one stereo file — a Suno or Udio export, a whole mix you want to rebalance — drop it in and Latent separates it into four stems (drums, bass, other, vocals) on your GPU. If you already have stems (your own instrumental plus an AI vocal, or several AI generations you want to layer), drop them all in at once and Latent skips separation entirely. The file count tells it which path you want: one file means separate, two or more means mix.

When separation helps — and when it doesn't

Separation is the on-ramp for whole-track exports: it's how you get from one stereo file to something you can rebalance. It's also genuinely useful for isolating an acapella or instrumental to use somewhere else. It's not a fix for a bad mix underneath — separation reveals what's there, it doesn't improve it. If your track already sounds thin or muddy, separating it won't change that; it just gives you the parts to work on.

The null-check badge — what "−38 dB residual" actually means

After separation, Latent measures the difference between the stem sum and your original track and shows it as a badge — something like "Stems reconstruct the original to −38 dB". This is a honesty feature no competitor offers: it proves the separation didn't eat your track. A more negative number is better (more of the original is accounted for). A number close to 0 would mean the stems don't add back up to what you started with — a sign of aggressive or lossy separation. In practice htdemucs residuals are well below −30 dB for clean sources.

Level-matched A/B: why it matters

Louder sounds better — always. If you compare your remix against the original at different volumes, the louder one wins, every time, regardless of what's actually better. Latent's A/B button matches loudness between your remix and the source before comparing, so you're judging the mix, not the volume. Use it whenever you're deciding whether a change is actually an improvement.

Export modes: mix / stems / instrumental / acapella — which to use

Mix — the full mixdown through the master bus. Use this when you're done and want one file.
Stems — each track post-insert, pre-master-bus, zipped together. Use this to hand off to another tool or archive the parts.
Instrumental — the mix with vocals muted. Exact, because mixing is linear — no re-render, no quality loss.
Acapella — only the vocals, soloed. Same exactness. Use this for distribution stores that need a clean vocal, or a remix base.

All exports are WAV — 16, 24, or 32-bit float, at native rate or 44.1/48 kHz. TPDF dither is applied automatically when reducing bit depth. FLAC and MP3 export are coming in a later phase.

Browser and hardware requirements

Stem separation runs on your GPU via WebGPU. You'll need a WebGPU-capable browser — Chrome or Edge recommended. Firefox and Safari don't yet ship WebGPU everywhere; if your browser can't run it, Latent will tell you exactly what's needed rather than silently failing. The first separation in a session includes a one-time shader compile of about a minute — after that it's fast. On the web, track length is capped at 15 minutes for memory reasons. A desktop app with native GPU and uncapped length is planned for later.

A short glossary

Stem: A single isolated component of a track — vocals, drums, bass, or "other" (everything else). Mixing stems means rebalancing these parts against each other.
WebGPU: The browser API that lets Latent run AI separation on your graphics card instead of a server. It's why nothing has to upload.
htdemucs: Meta's open Demucs separation model, ported to Rust. The thing that actually splits your track. ~84 MB, downloaded once and cached.
M/S: Mid/Side — a way of working with stereo as the center (Mid) and the stereo difference (Side). Useful for keeping bass mono or widening highs.
LUFS: Loudness measured the way streaming platforms measure it — not the same as peak volume. Latent can normalize your export to a target (Spotify −14, Apple −16, YouTube −14, Club −9).
True Peak: The real peak level after accounting for inter-sample peaks a simple sample reading can miss. Too high and you get distortion after lossy encoding; the master-bus limiter keeps it at −1 dBTP by default.