Separation Strategy¶
Flags: --separate-missing, --reseparate, --separator
Suno provides stems, but sometimes they're incomplete or low quality. AI stem separation (Demucs, UVR) can generate alternatives from the full mix.
Decisions involved¶
- Run separation at all? Only needed if stems are missing, damaged, or have too much bleed. The automatic pipeline skips this unless you pass
--separate-missing. - Demucs vs UVR? Demucs (Meta's model) is the default — good general-purpose separation. UVR (Ultimate Vocal Remover) can be better for vocal isolation specifically.
- Targeted re-separation? Use
--reseparateto re-run separation on the full mix or a specific stem (e.g. to isolate vocals from a bleed-heavy stem). - Lighter/faster model vs slower/better? Demucs
htdemucsis the default balance. Heavier models take longer but produce cleaner stems.
When to use each flag¶
| Flag | Use case |
|---|---|
--separate-missing |
First-time processing — generate stems that Suno didn't provide |
--reseparate |
Re-run separation after changing settings or to try a different backend |
--separator demucs |
General-purpose separation (default) |
--separator uvr |
When you specifically need cleaner vocal isolation |
Example¶
Combining with stem quality judgment¶
Run --choose-stems after separation to automatically compare originals vs generated stems: