Audio Stem Mixer 🎚️

Audio is ComfyUI's unloved stepchild. You can generate a vocal stem, an instrumental, a drum line - but stitching them into one listenable track usually means exporting, firing up a real DAW, and going back and forth. AudioStemMixer from the ComfyUI-Stem-Mixer pack is a genuinely polished answer to that: a full multitrack mixer that lives inside the node graph, with faders, pan, mute/solo, a 3-band EQ per track, eight effects, VU meters, and a real-time waveform preview. You mix with your ears, hit Run, and get a standard AUDIO output ready to save. For a niche pack with no community lore behind it yet, the README is unusually professional - and the code backs it up.

How it works

Here's the mental-model shift that trips everyone up: the node has no audio input ports. You don't wire an audio-generating node into it. Instead, you load files inside the node - a 📂 Load button per track, or just drag and drop one or more WAV/FLAC/MP3/OGG/M4A/AIFF files straight onto it. They get copied into ComfyUI/input/stem_mixer/ and referenced by path in the workflow JSON.

All that mixer state - tracks, gains, pan, EQ, loop region - lives in a single hidden state string widget that the frontend writes behind the scenes. Your fader tweaks are just edits to that JSON, which is why a saved workflow reopens with everything intact.

When the workflow runs, mix() parses that state, loads each non-muted stem with torchaudio (forcing stereo and resampling everything to a common sample rate), then applies per-track EQ, the effects chain, gain, equal-power panning, and mute/solo logic - if any track is soloed, only the soloed ones contribute. Stems are padded to the longest and summed, the master fader is applied, and a soft-limiter catches peaks above 0.95 dBFS instead of hard-clipping. The preview you hear in the browser (a WaveSurfer UI) matches the rendered mix closely - the Python EQ even scales its gain by 0.7 to line up with the Web Audio preview.

One detail worth knowing: the node hashes its state plus the mtime of every referenced file, so moving a fader or swapping a file on disk triggers a re-render. No stale mixes.

Inputs and outputs

The info_schema is delightfully small because the UI does the heavy lifting:

state (STRING, multiline, default []) - the only input, and you never touch it directly. It's the serialized mixer session the JS writes.
audio (AUDIO) - the mixed result, a standard ComfyUI AUDIO dict. Wire it into PreviewAudio to listen, SaveAudio to write a file, or any downstream audio node.

That's it. The complexity is in the widget, not the ports.

Installing it

The painless part: there are no extra Python dependencies (torch and torchaudio already ship with ComfyUI) and no model downloads. Install via ComfyUI Manager (search "ComfyUI-Stem-Mixer"), or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/tritant/ComfyUI_Stem_Mixer

Restart ComfyUI and you'll find Audio Stem Mixer 🎚️ under audio/mixer. That's the whole install story - a rare relief in a world of dependency hell.

Where people get burned

The files don't travel with the workflow. Share the .json and the recipient opens an empty mixer, because the audio lives in their input/stem_mixer/ folder. You're sharing a recipe, not the ingredients.
Silence isn't a crash. If a referenced file is missing on execution, the node logs [StemMixer] Failed to load and skips it; if nothing loads, you get a silent audio output. Watch the console when a mix comes back empty.
It's load-from-disk only. If you hoped to wire in an in-memory AUDIO-producing node, you can't - pair this with anything that writes stems to disk instead.
Uploads are capped at 500 MB per file, and solo is exclusive (one soloed track at a time).

Use it when you've already separated your stems and want to batch-render a mix without leaving ComfyUI. For that job, nothing else in the ecosystem does it half as well.