Egystormy Models
ONNX exports of Meta's Demucs music source separation models, published by the Egystormy project. Free for everyone.
Models
demucs/htdemucs_fwd.onnx (+ demucs/htdemucs_fwd.onnx.data)
HTDemucs 4-stem ONNX export — separates audio into:
- vocals
- drums
- bass
- other
This is an ONNX export by Intel (from Intel/demucs-openvino) of Meta's Demucs model. The graph lives in htdemucs_fwd.onnx and its weights in the external data file htdemucs_fwd.onnx.data — download both and keep them together.
demucs/htdemucs_6s.onnx
HTDemucs 6-stem ONNX export — separates audio into:
- drums
- bass
- other
- vocals
- guitar
- piano
Produced by the Egystormy project from the official Demucs checkpoint, and verified numerically against PyTorch (per-stem cosine >= 0.9999998).
License
Demucs code and weights are MIT (Meta AI). The 4-stem files above come from Intel's ONNX/OpenVINO export of that same MIT-licensed model; the 6-stem export was produced by the Egystormy project from the official Demucs checkpoint.
Free for everyone.