Load Audio Separation Model

Every audio-reactive workflow in ComfyUI_Yvann-Nodes that needs to isolate a stem starts with this node. It doesn't do any analysis itself - it just loads (or downloads) the source-separation model that Audio Analysis and Audio Remixer will use to split your track into drums, bass, vocals, and "everything else."

What you're choosing between

The dropdown offers two models:

• Hybrid Demucs - the one you'll actually use. It's torchaudio's HDEMUCS_HIGH_MUSDB_PLUS pipeline, a hybrid transformer/convolutional separator trained on MUSDB. Four stems out: bass, drums, vocals, others. The tooltip calls it "fast, accurate," and it earns both adjectives - it's the modern default for this kind of work.
• Open-Unmix - the alternative, specifically the UMXL model from the sigsep/open-unmix-pytorch repo. Older and heavier per run, but a classic research-grade separator that some people prefer for the character of its vocal extraction. If Hybrid Demucs gives you weird artifacts on a particular track, swapping to Open-Unmix is a one-click experiment worth making.

How the loading works

There's no model inference on this node. Selecting Hybrid Demucs pulls the pretrained bundle from torchaudio's hub on first use and caches it, so the second workflow runs instantly. Selecting Open-Unmix downloads the umxl.pth weights through torch.hub and saves them to ComfyUI/models/openunmix/umxl.pth.

Here's the small trap: the README says models are downloaded to ComfyUI/models/audio_separation_model/, but the code actually writes Open-Unmix to ComfyUI/models/openunmix/. If you go looking for the file in the folder the README names, you won't find it. Not a big deal - the node finds it automatically - just don't let it send you down a rabbit hole.

The one output

audio_sep_model - a single connection that goes into Audio Analysis (for per-frame weight extraction) or Audio Remixer (for per-stem volume control). The type is the pack's custom AUDIO_SEPARATION_MODEL, so you can't accidentally wire it into a plain model slot.

Install and gotchas

Same install as the rest of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/yvann-ba/ComfyUI_Yvann-Nodes

or ComfyUI Manager → search ComfyUI_Yvann-Nodes → install → restart. The requirements pull in torchaudio and openunmix (plus the librosa/soundfile stack that comes with it), so the first install is noticeably heavier than a typical node pack.

Two things to know before you queue:

1. First run downloads weights. Give the first queue time - Hybrid Demucs fetches from torchaudio's hub, Open-Unmix from HuggingFace. If the Open-Unmix download fails (torch.hub is sensitive to flaky connections), you'll see a "Failed to download model" line in the console and the node returns nothing; re-running usually fixes it once the network cooperates.
2. It uses your GPU during separation. The model is loaded via ComfyUI's model management, so it lands on the same device you sample on. Tiny cards will feel it during the separation pass, though the analysis itself is quick.

This is a "load once, wire everywhere" node - you only need one instance in a graph, and Audio Analysis and Audio Remixer can both consume the same audio_sep_model connection.