MiniMaxRef Save Audio

MiniMax H3's whole headline trick is native, jointly-generated audio - dialogue and sound effects come out of the same forward pass as the video instead of being bolted on afterward. Which is great, until you realize you've got audio living inside a latent and no clean way to get it onto disk. MiniMaxRefSaveAudio is that way: it takes an AUDIO input and writes it to ComfyUI/output with the filename prefix you give it.

It's an output node, so it terminates that branch. The inputs are the file name, the format, and the audio. filename_prefix is a string, default Tenz/audio (the author ships as Tenz-8; the slash just makes a subfolder). format is where this node is nicer than the stock save: it's a dynamic combo that adapts to the format you pick - flac with no options, mp3 with a quality choice (V0, 128k, 320k), or opus with bitrates from 64k to 320k. So the format dropdown literally reconfigures itself based on what container you want, which is the small kind of UX that makes a pack feel designed rather than assembled. The audio input is the AUDIO you're saving, and the audio output passes it through unchanged so you can keep the signal flowing to other nodes.

Mechanically it's a thin wrapper around ComfyUI's audio-save helper - the same helper the stock node uses - which means files land in ComfyUI/output with the standard counter suffixing when a name collides. Like its image sibling, it forces re-execution every loop iteration, so a per-segment filename built from a loop index via MiniMaxRefJoinString gets honored on every pass instead of collapsing into one file. In the H3 workflow, that's exactly the pattern you want: each segment's audio saved as segment_002.mp3 rather than the loop reusing the first name.

What's not this node's job: mixing, loudness, or assembling a full soundtrack. The pack has other machinery for those - the Guide node can cut a per-segment clip from a synthesized master track, and Merge Videos From Paths will lay an AUDIO input over the final stitched film. This node is the end of the pipe: audio in, file out.

Install with the whole pack - ComfyUI Manager (search "MiniMaxRefDirector-ComfyUI") or

cd ComfyUI/custom_nodes
git clone https://github.com/eaglering/MiniMaxRefDirector-ComfyUI
cd MiniMaxRefDirector-ComfyUI
pip install -r requirements.txt

then restart. The only dependency beyond ComfyUI's bundled set is PyAV (av>=12.0.0), which is genuinely load-bearing here - the pack uses PyAV to read audio from video for the reference-conditioning path, so on a fresh install the audio nodes are the reason that pip step exists. Restart after installing, hard-refresh the browser (Ctrl+F5) if you updated front-end files, and you're set. One expectation to set: this is a v3 pack with no community trail yet, so when something's off, the README and the example workflow are your docs - but a save node this simple tends to just work.