Audio Export

Audio Export is the sample node from a pack called ComfyUI-LikeSpiderAI-UI, and before you trust it with your music video, know what you're actually getting: a demo. The pack itself is a "declarative UI framework" - one LikeSpiderUINode base class that lets node authors describe inputs in a config dict instead of hand-writing INPUT_TYPES. Audio Export is the reference example that proves the pattern works. That doesn't make it useless, though. It does one thing the built-in ComfyUI nodes mostly don't: writes audio to .mp3, .wav, or .flac with a real bitrate selector, and it needs nothing heavier than what you already run.

Why you'd reach for it

ComfyUI core ships a SaveAudio node, but it's wav-or-bust. Once a workflow hands you an AUDIO connection - say, from an LTX-2 audio decode or any VHS audio node - you often want an mp3 you can email someone or a small flac for archiving, and bitrate matters. That's this node's entire job. It's an output node, so it sits at the end of the chain: audio in, file on disk. Since it passes the audio straight through, you can also drop it mid-graph, save a checkpoint copy, and keep feeding the chain.

How it works

The node receives the standard ComfyUI AUDIO dict (a waveform tensor plus sample_rate), converts it to numpy, and peak-normalizes if the samples exceed ±1.0 so you don't clip. It writes a temp wav with scipy, then:

• mp3 - shells out to ffmpeg with libmp3lame at your chosen bitrate
• flac - ffmpeg with the flac codec
• wav - just copies the temp wav (no ffmpeg needed)

Files land in ComfyUI/output/audio/, auto-numbered as track_00000.mp3, track_00001.mp3, and so on, so reruns never overwrite each other. The filename is derived from a glob of what's already there, which is clever until it isn't (more below).

The inputs that matter

Only four, and only two you'll actually touch:

• audio - the AUDIO input. Wire your generated audio here.
• format - mp3 (default), wav, or flac.
• bitrate - a dropdown of 64 / 128 / 192 / 256 / 320 kbps, defaulting to 192. Only means anything for mp3; flac is lossless so it's ignored.
• filename - base name (default track); the counter suffix is added for you.

The output is audio, the same AUDIO it received - handy if you want to tee the file and keep processing.

Install

It's a one-file pack with no requirements.txt and no models to download, so install is trivial. ComfyUI Manager: search ComfyUI-LikeSpiderAI-UI. Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/Pigidiy/ComfyUI-LikeSpiderAI-UI

Restart ComfyUI and look for Audio Export under the LikeSpiderAI category. It needs numpy, torch, and scipy (scipy is the only non-obvious one - you almost certainly have it if you've installed anything audio-adjacent) and ffmpeg on your PATH for mp3/flac. Windows users with a standard ffmpeg install are fine; if ffmpeg isn't a recognized command, only wav will work until you fix that.

Where people get burned

Read the source and the gotchas fall out:

• The sample rate is hardcoded to 44.1 kHz. The node ignores audio["sample_rate"] entirely and writes the temp wav at 44100. Feed it 48 kHz audio from a video model and your "mp3" plays slow and low - the data isn't resampled, just mislabeled. If your source is 44.1k, you'll never notice. Check your source.
• The temp file is always temp_input.wav. Two AudioExport nodes in the same queue can clobber each other's intermediate file mid-run. Don't run them in parallel.
• It's a demo with a one-commit repo (June 2025) and effectively zero community footprint. There's no roadmap, no issue tracker culture, and the README only exists to show off the framework. If it breaks on a ComfyUI update, you're on your own - the fallback is VHS's audio save node, which has the same ffmpeg dependency but way more maintenance behind it.

Verdict

For a beginner who just wants an mp3 of that LTX audio without installing a second pack, Audio Export is fine - it's small, installs clean, and the bitrate selector is genuinely nice. Treat it as what it is: a well-made sample node that happens to be useful. And if you ever write your own ComfyUI nodes, the UI_CONFIG pattern it demonstrates is actually a pleasant way to skip the INPUT_TYPES boilerplate.