Collected sources and patterns will appear here. Add from search, explore, or the patterns library.
List<SemanticToken> -> Audio<HighFidelityVocal>, Audio<HighFidelityAccompaniment>
Decode coarse semantic draft tokens generated in stage 1 through a stage 2 neural vocoder/refiner to produce high-fidelity multi-track audio stems.
Problem it solves
Generating high-fidelity multi-track audio in a single autoregressive pass is computationally expensive and prone to audio artifacts.
Consumes
Emits
The real projects this mechanism was found in. Attribution is the point — this is how the best teams actually do it.