HTDemucs β ONNX conversions
ONNX conversions of the Demucs v4 (Hybrid Transformer) family, packaged for
SonicDemixer (on-device stem separation for macOS). All
graphs share one contract: input mix [1, 2, segmentSamples] float32
planar waveform β output sources [1, numStems, 2, segmentSamples],
opset 17, STFT/iSTFT embedded in the graph (no complex dtypes).
| File | Stems | Notes |
|---|---|---|
htdemucs.onnx |
drums, bass, other, vocals | baseline 4-stem |
htdemucs_6s.onnx |
+ guitar, piano | 6-source variant |
htdemucs_ft.onnx |
drums, bass, other, vocals | the fine-tuned bag's four per-source models fused into ONE graph β each submodel contributes its specialty stem (mapping verified against the bag weights). Higher quality, ~4Γ the compute. |
Integrity: files are LFS-tracked; the LFS OID is the SHA-256 that consuming apps verify after download.
License & attribution
MIT. These are derivative conversions of:
- Demucs / HTDemucs β code and pretrained weights Β© Meta Platforms (FAIR), MIT. https://github.com/facebookresearch/demucs
- ONNX-exportable HTDemucs + conv-based STFT/iSTFT β Mixxx Development Team demucs fork (GSoC 2025, Anmol Mishra), MIT. https://github.com/mixxxdj/demucs
Conversion scripts: tools/model-conversion/ in the SonicDemixer repository.
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support