Open-Unmix · UMX / UMXHQ / UMXL
A reproducible separation baseline with one model per source.
SeparationOpen-source code
- Start here
- Choose the exact checkpoint first. UMXHQ is a useful public-data baseline; UMXL has different training access and weight terms.
- Training data
- UMX uses MUSDB18; UMXHQ uses MUSDB18-HQ; UMXL uses a private stems collection.
- Training rights
- MUSDB restrictions and private training access remain separate from the code. UMXL explicitly restricts commercial weight use.
- License & output terms
- Code: MIT. UMXL weights: CC BY-NC-SA 4.0; check other checkpoint terms.
- Limitations
- Bidirectional processing uses future context, so the standard model is not a causal realtime separator.
- Technique
- Bidirectional LSTM + spectrogram masking
- Runtime
- Local offline PyTorch inference and documented custom-data training.
- Parameters
- Not itemized for this release/family. Parameter sizes shown elsewhere may exclude encoders, decoders or other runtime components.
Sources reviewed . Training and rights notes summarize provider disclosures. A model license does not clear every source recording or generated output. Parameter counts refer to the named variant or component, not total runtime memory.