v0.1.0

First release.

DBS Soundtrack Separation for ComfyUI: separate a soundtrack into music, sound effects and speech with the official Dynamic Band Splitting models (DBS-TIGER-DnR, DBS-BSRNN-DnR), on CUDA or the CPU.

• Nodes: DBS Loader and DBS Separate (music / sfx / speech); channel mean (the official usage) or each channel (keeps stereo); any sample rate (ComfyUI's resampler); the stems add back up to the input.
• Workflows: the three stems as FLAC; a soundtrack without dialogue (music + sfx with ComfyUI's Merge Audio) plus the dialogue; the same for a video (Load Video → … → Save Video).
• Models: tools/download_model.py fetches both (23 MB) at pinned revisions and checks their SHA-256. The nodes never download.

Checked (details in docs/VERIFICATION.md):

• CPU: bit for bit the official package's README usage on the project page's film clips, 6 of 6 cases (both models, every overlap-add path, per channel).
• CUDA: the official code is itself not deterministic on the GPU; with PyTorch's deterministic algorithms on both sides, bit for bit 4 of 4 (30 s clips).
• Real ComfyUI queue on CPU and RTX 5090: the saved FLAC equals the official stems to the 16-bit rounding; the video workflow writes a video with the new soundtrack.
• Clean install from this repository into a fresh ComfyUI v0.38.0; CI on Ubuntu and Windows, ComfyUI v0.38.0 and v0.42.1.

Speed (measured here): RTX 5090 16x (DBS-TIGER) / 22x (DBS-BSRNN) real time; CPU with 24 threads 0.5x / 1.8x.

Licenses: pack code MIT; vendored DBS code Apache-2.0 (upstream commit ff52646); models Apache-2.0, trained on DnR v2 (CC BY 4.0).