Drums using MDX

Want the beat out of a track - for a remix, for a drum-n-bass video edit, or just to hear what the percussion is actually doing? Drums using MDX (AudioSeparateDrums) from the Audio Separation (Demix) pack pulls the drum stem out in a single pass, plus the everything-else Complement. No external tool, no audio app, no model hunting.

Mechanically it's the same MDX-Net pipeline as the pack's other single-stem nodes: the audio is resampled to the 44.1 kHz the models were trained on, mono is faked to stereo, and STFT spectrogram chunks run through a cross-transformer model before being stitched back together. The Complement output is just the input minus the drums - cheap, but the pack is upfront that a dedicated instrumental model usually beats a subtraction-based Complement if what you actually want is "everything except drums."

The inputs

• model - two choices, KUIELab A and KUIELab B (default A). A and B are alternate training runs of the same family; flip to B if A leaves too much bleed. ⬇️ marks models not yet downloaded; they auto-fetch to ComfyUI/models/audio/MDX on first run.
• segments - chunks processed per pass (1–64, default 1). Crank it if you hear discontinuities; it trades VRAM.
• target_device - cpu or cuda.
• input_sound - the AUDIO to demix, from core LoadAudio or another node's AUDIO output.

The outputs

Two standard AUDIO outputs, both at 44.1 kHz stereo:

• Drums - the separated percussion.
• Complement - the input minus the drums.

Wire either into core PreviewAudio / SaveAudio. Because both stems share sample rate and channel count, they remix cleanly if you want to e.g. sidechain the rest of the track against the extracted beat.

Installing it

Same pack as all the others, so one install gets you all six nodes. ComfyUI Manager → search Audio Separation, or:

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

Only new dependency is seconohe; the rest (torchaudio, numpy, safetensors, tqdm) ships with ComfyUI. The models are the smallest in the pack - around 21 MB - so first run is nearly instant after download.

When to use it

Use it when the drum stem alone is the point. If you're pulling bass and drums and vocals for a real remix, stop running three MDX nodes and use the Demucs Audio Separator - four stems, one pass, better quality. The trade: Demucs models are hundreds of MB and slow. This node is the fast, surgical option, and for a single stem it's usually plenty.

Standard pack gotcha: if the model dropdown desyncs after a download, press R in ComfyUI to refresh the list.