Music Source Separation with Band-Split RoPE Transformer
Paper • 2309.02612 • Published • 2
MVSep Mega 53 stems - KICK MLX
The 8 mask heads that KICK Studio uses (vocal, drums, bass, guitar, keys, synth, strings, wind) of the
MVSep Mega 53 stems BS-RoFormer (mvsep_mega_model_bs_roformer_53_stems_v1.ckpt), converted by KICK to safetensors
(fp16, 125.2 M parameters) for the Apple Silicon runtime of KICK Studio (native/mlx, the kick-mlx bridge: Swift + MLX, GPU).
The weights are unchanged; the other 45 heads are dropped and tensor names are mapped.
This is not a generic MLX model: key names and roformer_config.json follow the kick-mlx loader.
Files
mvsep-mega-stems-kick-mlx.zipcontains the foldermega-stems-mlx/:roformer.safetensors- 2411 tensors, fp16.roformer_config.json- architecture (dim256,depth12,heads8,num_stems8, band layout) and the stem order.
- Output: a mask
[8, T, 4100]per chunk (8 s, STFT 2048 / hop 512), the same layout as the ONNX export used on Windows.
Accuracy and speed
Against the PyTorch reference (MSST code) on a real STFT segment of 690 frames, the masks match with a relative error of 0.12 %. MacBook Air M4: about 2.2 s per 8 s segment on the GPU.
Credits and license
- Weights, config and model code: Roman Solovyev (ZFTurbo) -
ZFTurbo/Music-Source-Separation-Training (code: MIT;
checkpoint released at tag
v1.0.21, no separate license published for the weights). - Architecture: BS-RoFormer (Lu et al., arXiv:2309.02612), implementation by lucidrains/BS-RoFormer (MIT).
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support