BS-RoFormer HyperACE (music source separation)
This model page is essentially a checkpoint drop with evaluation numbers: pcunwa/BS-Roformer-HyperACE publishes updated “v2” weights for both instrumental and vocals stems, with links to MVSEP Quality Checker entries and a set of reported metrics (SDR, SI-SDR, etc.). The author claims the instrumental v2 weights hit a top score on at least one MVSEP metric for the Multisong dataset.
The key caveat is that it’s not a turnkey inference package. The model card’s usage guidance is basically “replace the model file with one from this repository,” and notes the weights build on anvuew/BS-RoFormer. So this entry is most useful if you already have a BS-RoFormer-based separation setup (or are benchmarking checkpoints) and want to try a newer set of weights without re-training. A good first test is to run a small batch of tracks through the same settings you use for your current baseline, then A/B stems for bleed and artifacting rather than relying only on headline metrics.
Quick stats from the listing feed: 9 likes · 564 downloads.
Source listing: https://huggingface.co/models?sort=modified