MiniMax H3 Audio Reroll (Experimental) / 音频重生成（实验性）

Here's the problem this node exists to solve. MiniMax H3 samples video and audio in one loop, so if you don't like the score, changing the soundscape text means regenerating the entire clip - and with it the whole visual lottery: framing, performance, lip-sync. There's no official "reuse this picture, new sound" workflow. Audio Reroll is an experimental attempt to build one on top of machinery H3 already has.

How it works

H3 can't freeze half of a jointly-denoised pack mid-sampling, but it does have reference conditioning - the same mechanism used for keyframes and image references. The node takes your finished video latent and hands it back to the DiT as a whole-clip video reference with video_cond_strength at 1.0. At that setting the reference enters completely un-noised and the model treats it as "already clean, take it as given." The audio stream starts from zeros (if reset_audio_latent is on), so the sampler writes a genuinely new track against a strongly-conditioned picture.

Inputs: conditioning (from the same H3 conditioning node as the original run, with the soundscape/music text changed), av_latent (the finished run whose picture you want), video_cond_strength (default 1.0 - lower lets the picture drift), and reset_audio_latent (default on; off biases toward the original track).

Outputs: positive conditioning, the new latent, video_passthrough (the original video stream, bit-identical), and an info string.

The fine print - read this before you queue

The pack's own measured findings are refreshingly honest, and they matter:

• The picture is strongly preserved, not frozen. The test measured 43.6 dB PSNR between reroll and original - very close, on the order of a lossy re-encode. Good, but not identical.
• It's slower than generating from scratch. Feeding the whole clip back as a reference doubles the sequence length, and attention is quadratic. Measured: 26 s/it for the original run, 63 s/it for the reroll - about 2.4×. The packed reference tokens ride along on every sampling step.
• It requires the ref2va checkpoint path. The fl2va conditioning node has no minimax_refs path, so this node won't work without ref2va weights.
• Validation is one clip, one seed. The preservation numbers describe that one experiment; your mileage is genuinely unmeasured.

So the honest use case is when identity matters more than time: you nailed the performance and framing, the dialogue and lip-sync are right, and you only want a different ambience or score underneath. If you just want a cheap audio variation, this is the wrong tool - regenerate at a fixed seed, or skip the DiT entirely.

The bit-identical alternative

If you need the original frames byte-for-byte, don't re-sample at all. Take the video_passthrough output, decode it, and mux the new audio over it downstream - that's one VAE decode and one audio decode, seconds instead of minutes. The trade-off: the new audio was generated without seeing the video, so it won't lip-sync to existing dialogue. Use it for ambience and score, not for speech.

Install

Part of the MiniMax-H3 Tools pack (ComfyUI 0.30.0+). Search "MiniMax H3 Tools" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Rinne414/ComfyUI-MiniMaxH3-Tools.git

Restart ComfyUI, find MiniMax H3 Audio Reroll (Experimental) under MiniMax H3. No extra dependencies for the node itself - but it does nothing until you have a full H3 stack loaded (model, text encoder, both VAEs), and the ref2va weights specifically.