---
license: apache-2.0
tags:
  - audio
  - source-separation
  - demucs
  - onnx
  - onnxruntime
library_name: onnx
---

# HTDemucs (ONNX Runtime) — Gridshift mirror

This is a mirror of [gentij/htdemucs-ort](https://huggingface.co/gentij/htdemucs-ort),
re-hosted under the [Gridshift Studio](https://huggingface.co/gridshiftstudio) organization
so that the desktop app's model-distribution pipeline remains stable across
upstream changes.

**Weights, format, and license are unchanged from upstream.**

## Usage

Consumed by [Gridshift](https://gridshift.studio)'s stem-separation feature via
the Rust stem-splitter crate (`reqwest` + `onnxruntime`). The `manifest.json`
in this repo lists the ONNX model file with its SHA-256 and size for
client-side integrity verification.

- **Audio format:** 44.1 kHz stereo, 343 980 samples per inference window
- **Stems:** drums, bass, other, vocals
- **Format:** ONNX Runtime `.ort` (converted from PyTorch HTDemucs)
- **File size:** ~200 MB

## License and attribution

Apache-2.0, inherited from the upstream Demucs project by Facebook Research.
The conversion to ONNX Runtime was done by [gentij](https://huggingface.co/gentij);
we redistribute that conversion without modification.

Please cite upstream Demucs if you use this model:

```
Défossez et al., "Hybrid Transformers for Music Source Separation", 2022.
```