Bass using MDX

Sometimes you don't want vocals, instruments or drums - you want the low end. Bass using MDX (AudioSeparateBass) from the Audio Separation (Demix) pack isolates the bass stem and gives you everything-else as a bonus, in one run.

It's the focused sibling of Vocals using MDX: same MDX-Net machinery (spectrogram chunks through a cross-transformer model, resampled to the 44.1 kHz training rate, mono faked to stereo, batches handled one track at a time), just weights trained for bass. Where the Demucs node in this pack hands you four stems at once, this one is for when you know exactly which stem you're after and you want it cheap and fast.

The inputs

• model - just two choices: KUIELab A and KUIELab B (default A). Both do the same job; A vs B are different training runs of the same idea. Try the other if one sounds mushy. ⬇️ means not downloaded; models auto-fetch to ComfyUI/models/audio/MDX on first run.
• segments - chunks processed per pass (1–64, default 1). More = more VRAM, fewer seam artifacts.
• target_device - cpu or cuda.
• input_sound - the AUDIO to demix, e.g. from core LoadAudio.

The outputs

Two standard AUDIO outputs:

• Bass - the separated bass track.
• Complement - the input minus the bass.

Both are 44.1 kHz stereo, so they line right back up if you want to remix or duck the bass against the rest.

Installing it

It's part of the AudioSeparation pack, so it installs with the rest. ComfyUI Manager → search Audio Separation, or:

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

The only new dependency is seconohe; torchaudio, numpy, safetensors and tqdm are already in ComfyUI. MDX models are tiny (21–65 MB), so the first download is quick.

When to use it

This is a workflow-tail node: you're building a remix, an audio-reactive visual where you want bass energy to drive the low-frequency band, or a stem for a DJ set, and you don't need the other three stems cluttering the graph. If you find yourself wanting bass and drums and vocals in the same project, skip the three MDX nodes and grab the Demucs Audio Separator instead - it outputs all four stems in one pass. But for the single stem, this node is exactly as much tool as you need.

Same refresh gotcha as the whole pack: if the model list looks stale after a download, press R in ComfyUI to rescan.