MiniMax H3 Audio Mix (T8)

H3 hands you two audio streams most of the time: the source track you fed in and the audio the joint transformer generated around it. Neither alone is usually what you want to ship - the source is clean but doesn't breathe with the new footage, the generated audio is alive but can drift off-mic. MiniMaxH3AudioMixT8 is the in-graph mixdown: it takes both, balances them, ducks one under the other, and hands back a single mixed_audio you can drop straight into a save or preview node.

This is the node you reach for after a lock_source or remix_source run, when the Conditioning node's mux_audio isn't quite the blend you want. It's a small, honest utility - resample, gain, duck, limit, out - but it's the difference between shipping a clip where the dialogue fights the room tone and shipping one where they sit together.

What the knobs mean

• source_audio / generated_audio - the two inputs. Order matters for the ducking math: the source drives the duck envelope.
• source_gain_db (0) / generated_gain_db (−6) - your starting balance. The defaults already favor the source by 6dB, which is the right instinct for dialogue-first H3 output.
• duck_generated (0.5) - how much the generated track gets pulled down whenever the source is loud. 0 = no ducking (both play flat), 1 = heavy ducking. For a voice-over on top of generated ambience, 0.4–0.6 is the sweet spot.
• output_sample_rate (source) - pick source, generated, or an explicit 48000/44100/32000. H3's native audio is 32kHz; if your deliverable is a video at 48k, set it here rather than resampling later.
• peak_limit_dbfs (−1) - a brickwall ceiling so the blend never clips. −1dBFS is a sensible broadcast-ish default; the min is −12 if you want serious headroom.

Under the hood it's the classic chain: resample both streams to the target rate, match channels, pad to equal length, apply the dB gains, derive a ducking envelope from the source's amplitude, and run the result through the limiter. No AI involved - pure, predictable DSP, which is exactly what you want at the end of a workflow.

Wiring

Two AUDIO in, one mixed_audio out. The natural source for the second input is the generated_audio output of MiniMaxH3AVDecodeT8, and the natural source for the first is your original track or the source_audio pass-through from MiniMaxH3AudioLatentControlT8. Wire the mix into your video save node's audio input and you're done.

One honest caveat: it won't fix a bad generated track. If the generated audio has a voice that wandered off, ducking won't bring it back - that's what the Perceptual Drift Audit is for, and the answer there is usually re-running with lock_source.

Install

Same pack, same routine:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8

Restart ComfyUI. It's in the T8/MiniMax H3/Audio menu, needs no dependencies and no models - pure tensor audio math on the CPU/GPU whichever ComfyUI gives it. The pack's 02-audio-control workflows show it wired after a remix, which is the fastest way to hear what the defaults do to your own material.