Audio Remixer

Most of the pack is about turning audio into visuals. Audio Remixer is the odd one out: it turns audio into different audio. It splits your track into four stems, lets you boost or mute each one, and merges them back into a single file. It's a karaoke machine, a bass-booster, and a "remove the drums so the generator keys off the vocals" tool, all in one node.

The mechanism

It runs the same separation model you use for analysis - that's why it needs the audio_sep_model from Load Audio Separation Model - to split the track into bass, drums, vocals, and others. Then it applies a per-stem volume and recombines them.

The four volume knobs (drums_volume, vocals_volume, bass_volume, others_volume) all run from -10 to +10:

• -10 = mute (the stem is dropped entirely)
• 0 = the original level
• +10 = a healthy boost (the code scales it up to 10×)

So vocals_volume = -10 gives you an instrumental; bass_volume = 6 gives you a bass-pumped version of the same track; others_volume = -10, vocals_volume = 0 isolates the voice against sparse backing. The output merged_audio is the full mixed track at the model's sample rate, ready to feed PreviewAudio or save out.

Why you'd actually use it

The honest use case inside a reactive-animation workflow: you can remix the audio before it hits Audio Analysis, so the visual responds to the mix you want rather than the mix the artist shipped. Isolate the kick drum by muting everything else, and the weight curve becomes pure rhythm. Or strip the vocals so the analysis keys off the instruments and the animation isn't chasing the singer.

Outside the animation context it's a genuinely handy utility - a quick instrumental for a demo video, a louder bass for testing. It's not going to replace a real DAW, and the separation isn't perfect (vocals bleed into "others" on dense tracks), but for a node that costs you one wire, it's a lot of mixer.

Gotchas

• The first run downloads the model weights, same as everywhere else in this pack - Hybrid Demucs from torchaudio's hub, Open-Unmix from HuggingFace. Give it a minute.
• Model artifacts show up in the "others" stem, and some artifacts land in vocals. If the remix sounds phasey, that's separation quality, not your settings - try Hybrid Demucs if you started on Open-Unmix (or vice versa), since they artifact differently.
• The output sample rate is the model's rate (the separation model resamples internally), so don't be surprised if the merged file isn't at your source's exact rate. For driving Audio Analysis downstream it doesn't matter - that node resamples again.

Install

Same pack, same drill:

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

or ComfyUI Manager → search ComfyUI_Yvann-Nodes → install → restart. And remember it shares the AUDIO_SEPARATION_MODEL connection - you can hang both Audio Remixer and Audio Analysis off a single Load Audio Separation Model node without loading the model twice.