Vocals and texture. The instruments come from the engine.
bpm and key are a request, not a lock — the model drifts off both. The align pass is what actually puts it on the grid.
Drop the session.json from a stem export. The take is stretched onto that song’s bars, and optionally moved into its tuning — which may not be 12-TET, and that is the point.
session.json
Strength 1 puts every note exactly on a degree, which stops sounding like a person. Moves bigger than the limit are left alone — a note that wants 700 cents is a tracking failure, not a flat note.