Download Open Unmix Model ⚡🅡🅞🅣🅘

This node's entire job is fetching a pretrained stem-separation model so the rest of RyanOnTheInside's audio-separation nodes have something to run. Open-Unmix (UMX) isn't Ryan's own model - it's a well-established open-source music source separation project (from the Sigsep research group), and this node is a convenience wrapper: pick a variant, it downloads the checkpoint, and hands you an object the pack's separation nodes can consume.

If you've ever used a vocal remover or stem-splitter tool anywhere, this is that same category of model - split a mixed track into vocals, drums, bass, and "other."

How it works

You choose one of two model variants, the node downloads (or loads, if already cached) the corresponding pretrained weights, and outputs an OPEN_UNMIX_MODEL object. That's it - no separation happens here, this node is purely the loader. Wherever you actually split a track is a different node further down the pack's audio-separation family.

The inputs and outputs that matter

• model_name (required, default umxl) - an enum with two choices: umxl and umxhq. umxhq is trained on the higher-quality MUSDB18-HQ dataset; umxl ("large") is trained on a substantially larger dataset and is generally the stronger default for real-world tracks, which is presumably why it's the default here.
• Output - OPEN_UNMIX_MODEL, a single typed object with exactly one purpose: feeding into this pack's stem-separation node.

How to install it

Via ComfyUI Manager: search "RyanOnTheInside," install, restart. Manual install:

cd ComfyUI/custom_nodes
git clone https://github.com/ryanontheinside/ComfyUI_RyanOnTheInside.git
cd ComfyUI_RyanOnTheInside
pip install -r requirements.txt

then restart ComfyUI. This specific node needs internet access on first use - it's downloading real pretrained weights, not shipping them in the repo - so the first run will pause while it fetches whichever variant you picked. After that first download, it caches locally and subsequent runs are instant.

Common issues & troubleshooting

First run hangs or times out. That's the model download, not a crash. Open-Unmix checkpoints aren't huge by diffusion-model standards, but on a slow connection or a sandboxed/offline environment (a cloud GPU with locked-down networking, for instance) the download can stall entirely rather than just being slow. If you're running somewhere without general internet egress, this node will fail outright - it has no offline fallback.

Switching model_name after the fact re-downloads. umxl and umxhq are separate checkpoints, not variants of the same file, so changing the dropdown triggers a fresh download rather than reusing what you already fetched. Not a bug, just worth knowing before you assume something's broken because a "quick" model swap suddenly takes a minute.

Separation quality is disappointing on a busy mix. This is a limitation of Open-Unmix itself rather than this loader node - it's a solid, well-tested separator, but it's not the newest or highest-fidelity stem-splitting model on the market. If you need cleaner separation on dense mixes, that's a "try the other variant" or "accept the ceiling of this model family" situation, not something this node's settings can fix.