MVSep Mega 53 Stems — guitar head (float16 safetensors)
The guitar head of ZFTurbo's MVSep Mega 53 Stems BS-RoFormer, converted to float16 safetensors for LumaBeat Monitor. The Monitor runs it on Apple silicon with MLX Swift after BS-RoFormer-SW: where Mega 53 hears guitar in BS-RoFormer-SW's "other" stem, that part moves to the guitar stem.
- Stem: guitar (one of the original model's 53; the shared transformer is complete, only the other stems' mask estimators are left out, so this stem separates exactly as in the full model)
- Audio: 44.1 kHz stereo; STFT n_fft 2048, hop 512
- Model: BS-RoFormer, dim 256, depth 12, 8 heads of 64, 62 bands, mask estimator depth 2 (MLP ×2)
- Inference: chunks of 882,000 samples (20 s), 2 overlaps
Files
| File | Bytes | SHA-256 |
|---|---|---|
mvsep_mega53_guitar.safetensors | 77,455,608 | 24fbd2bd29f134f94cb2396e5f43d6b8dcbcef621ae75783275b1c9c281c6fec |
mvsep_mega53_guitar_config.json | 1,080 | 3691ae04991564ef44fa93b9793e95e07d2e21e1df373e129fcc3f871a843306 |
Origin
A format conversion only: no training or fine-tuning was done here.
- Source:
mvsep_mega_model_bs_roformer_53_stems_v1.ckpt(SHA-256c62820893bbf86d4e734f966bd142d9157cfc8bb8e79e9d8f9ea553f3ff3519f), release v1.0.21 of ZFTurbo/Music-Source-Separation-Training (2026-04-20). - Converted with
torch.load(weights_only=True)(no code in the checkpoint runs), keeping theguitarmask estimator. The checkpoint stores its rotary frequencies rounded to float16; the Monitor recomputes the default ones. - Checked against ZFTurbo's PyTorch code on 24 s of music: float16 output 51.6 dB SDR from the reference.
License
The weights are published in an MIT-licensed repository (Copyright (c) 2024 Roman Solovyev (ZFTurbo)), with no separate statement for them; this repository follows that. Architecture: BS-RoFormer (Lu et al., 2023), code by lucidrains and ZFTurbo (both MIT).
If you hold rights to these weights and want them credited differently or taken down, please open a discussion on this repository.