Various using MDX

Every node in the Audio Separation pack has a normal job - vocals, instruments, bass, drums. And then there's Various using MDX (AudioSeparateVarious), the misc drawer. Its model list is just three deep: FoxJoy Reverb HQ, KUIELab Other A, and KUIELab Other B. If a stem doesn't fit the other nodes, this is where you look.

What do you actually use it for? Two things. The KUIELab "Other" models separate the leftover noise, artifacts and ambience that isn't voice, music, bass or drums - the stuff that keeps a vocal isolate from sounding clean. The FoxJoy Reverb HQ model is the interesting one: it extracts the wet reverb from a track. That's a genuinely handy trick. Sample the reverb tail from a song you like and you can synthesize atmosphere, or duck the extracted reverb to make space in a remix. Reverb extraction is niche, and this is one of the few places you get it inside ComfyUI.

The mechanism

Same MDX-Net core as the pack's other MDX nodes - STFT chunks through a CrossTransformer model, resampled to 44.1 kHz, mono faked to stereo, batches processed one track at a time. The one structural difference is cosmetic: where the other single-stem nodes label outputs Stem/Complement, this one names them Main and Complement, because its model choices have two different primary stems ("Other" and "Reverb"). Mechanically it's the same input-minus-primary complement.

The inputs

• model - three choices, ⬇️/💾 status icons, auto-downloaded to ComfyUI/models/audio/MDX on first use. Start with FoxJoy Reverb HQ if you want reverb, either KUIELab for "other."
• segments - 1–64, default 1; more chunks per pass = more VRAM, fewer seams.
• target_device - cpu or cuda.
• input_sound - any AUDIO from core LoadAudio (or another node's AUDIO output).

Installing it

It's in the same pack, so install once and all six nodes appear. ComfyUI Manager → search Audio Separation, or:

cd ComfyUI/custom_nodes/
git clone https://github.com/set-soft/AudioSeparation
pip install seconohe
# restart ComfyUI

Dependencies are minimal - seconohe plus things ComfyUI already ships (torchaudio, numpy, safetensors, tqdm). Models are small and download fast.

Honest notes

This is the node you reach for when nothing else fits, not the one you build a workflow around. Expect the "Other" results to be messier than the vocals/instrumental stems - "everything that doesn't fit" is a hard thing to train for, and it shows. The reverb model is the sleeper value here; the "other" models are more of a curiosity unless you're doing forensic cleanup of a bad source.

Same refresh gotcha as the rest of the pack: if the dropdown desyncs after a download, hit R to rescan ComfyUI. And remember the output is AUDIO type, so it feeds core PreviewAudio / SaveAudio (or a remix workflow) directly.