silverdaw/htdemucs-ft-onnx

Silverdaw Backup Stem Pack — htdemucs fine-tuned (ONNX)

An ONNX export of the Hybrid Transformer Demucs fine-tuned 4-stem model (htdemucs_ft), used by Silverdaw as its backup stem-separation engine. htdemucs_ft is a "bag" of four specialist models — one per source — so vocals / drums / bass / other are all native outputs.

Silverdaw's primary engine is the higher-quality RoFormer pack pair (vocals + rhythm); this model is used when those packs are not installed, or when the user explicitly selects the backup engine.

This repository is a re-host of StemSplitio/htdemucs-ft-onnx so Silverdaw can serve every model it needs from one self-owned location. The weights and the export are unchanged (byte-identical; same SHA-256).

Files

File	Size	Stem	SHA-256
htdemucs_ft_vocals.onnx	~302 MB	vocals	8c5d5e2da1f27050240bb80236673307ee3b40d4b064066d9350f4d64bfd544d
htdemucs_ft_drums.onnx	~302 MB	drums	f76b68af36066e38885b369299b5032a861038f9b49da5aa6cf1c31cfa69cf27
htdemucs_ft_bass.onnx	~302 MB	bass	2a74d9283fc2336fcc58d50f87a7080aff57aea372f65cfe3f0211ea1ff16182
htdemucs_ft_other.onnx	~302 MB	other	90e11806c1bb558ca9d9c7e909d28a2854f7f217982e90482dbed6442513daad

Each file is a self-contained .onnx (fp32 weights, no external .data). fp32 is chosen over the fp16 variants for the CPU execution provider, where fp16 ops are commonly emulated (slower) rather than accelerated.

Provenance & license

License: MIT.

• Weights — © Meta / Alexandre Défossez et al., Demucs v4 (Hybrid Transformer Demucs, fine-tuned), released under MIT.
• ONNX export — © StemSplitio (StemSplitio/htdemucs-ft-onnx).
• Architecture / paper — Hybrid Transformers for Music Source Separation (Rouard, Massa, Défossez; arXiv:2211.08553) / facebookresearch/demucs (MIT).

Intended use

Designed for offline (non-real-time) stem separation inside a desktop DAW. It is not a real-time audio-callback model — run it on a worker/background thread.