Instrumental using MDX

There are two ways to get an instrumental in the Audio Separation pack: run Vocals using MDX and take the Complement output, or use Instrumental using MDX (AudioSeparateInstrumental) and get a model that was actually trained to isolate instruments. The README is blunt about which is better - a dedicated instrumental model usually beats whatever falls out of input − vocals. This node is that dedicated model.

It's the same MDX-Net machinery as the other MDX nodes in the pack: spectrogram-based neural demixing, resampled to the 44.1 kHz training rate, mono faked to stereo, batch in → batch out. The difference is the weights - 15 instrumental-specialized models rather than vocal ones, all inherited from the UVR5 ecosystem. The default Kim Jensen model is a solid, fast choice.

One useful quirk of this node's model list: it includes the karaoke models, which behave slightly differently from regular instrumentals. Where a normal instrumental tries to strip all vocals, the karaoke models deliberately keep secondary/background vocals along with the music. If you're after a true sing-along backing track, that's the one you want; if you want the instruments dead clean, stick with the straight instrumental models.

The inputs

Identical structure to its siblings:

• model - 15 instrumental MDX models (Kim Jensen, UVR 187/292/HQ variants…). ⬇️ = not downloaded, 💾 = on disk; they auto-fetch to ComfyUI/models/audio/MDX on first run.
• segments - chunks processed per pass (1–64, default 1). Bump it if you hear seam artifacts; it costs VRAM, not much else.
• target_device - cpu or cuda.
• input_sound - the AUDIO to demix, e.g. from core LoadAudio.

The outputs

Two AUDIO outputs, both in ComfyUI's standard format:

• Instrumental - the separated music.
• Complement - the input minus the instrumental, i.e. the vocals. Same caveat in reverse as the vocals node: if you actually want a lead-vocal track, the dedicated vocals node will beat this Complement output.

Wire either into core PreviewAudio or SaveAudio and you're done. Both stems share the sample rate of the output, so they remix cleanly.

Installing it

It ships in the same pack as everything else - ComfyUI Manager (search Audio Separation), or manually:

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

Only extra dependency is seconohe; torchaudio, numpy, safetensors and tqdm are already in a ComfyUI environment. Models are small (21–65 MB), so the first-run download is over before your coffee cools.

The gotcha worth knowing

If the model dropdown looks stale after a download - say you picked a karaoke model, it downloaded, and the list shifted - press R in ComfyUI to refresh and re-select. Also: don't reach for this node when your target is vocals. It does one job, instruments, and the instrument-specific siblings in the same pack cover bass and drums when you want those on their own. For everything else, the Instrumental output here is the workhorse.