demucs-onnx — exported graphs

Demucs music source separation, exported to ONNX so it can run without PyTorch. These are the weights for demucs-onnx, which runs them from Python, C++, Rust, Java, C# and JavaScript.

Try them in the browser: the demucs-onnx Space.

What is in here

Each model is a pair — <name>.onnx, the graph, and <name>.json, a sidecar naming its stems, sample rate and STFT parameters. A bag (htdemucs_ft) is a manifest listing four member graphs and the per-stem weights they are blended with.

file	stems	notes
htdemucs	drums, bass, other, vocals	the v4 default
htdemucs_fp16	same	half-precision weights
htdemucs_int8	same	quantised; much faster on a CPU
htdemucs_6s	+ guitar, piano	six-stem model
htdemucs_ft	drums, bass, other, vocals	bag of four fine-tuned networks, best quality
hdemucs_mmi	drums, bass, other, vocals	the v3 hybrid model

Using them

pip install "git+https://github.com/HRSadeghi/demucs-onnx.git#subdirectory=python"
hf download HRSadeghi/demucs-onnx htdemucs.onnx htdemucs.json --local-dir models
demucs-onnx song.wav -m models/htdemucs -o separated

The graph is cut at the complex boundary: the STFT, the ISTFT and the overlap-add happen in host code, and only the conv encoder / transformer / conv decoder is in the ONNX graph. The project README explains why, and every language binding in the repository reads exactly these two files.

Licence

The runtime code is MIT. The weights are Meta's, and carry the terms of the Demucs licence — non-commercial (CC BY-NC 4.0) for the v3/v4 checkpoints. Check that before you put this in a product.