沐阳 H3 · 导演台音频设置

Director-style packs all hit the same wall eventually: a node that wants eleven settings can't have eleven widgets, because the frontend has to serialise them, the UI has to lay them out, and every new option breaks older saved workflows. So the options get packed into one object and shipped over a single socket. That's all H3AudioSettings is.

It's a transport node. It samples nothing, mixes nothing, and produces no waveform. It collects seven widgets into a MYANG_H3_AUDIO bundle that the Director consumes when it builds the audio half of a segment.

What's in the bundle

Two independent things ride on the same socket, and it's worth separating them in your head before you touch anything.

The audio refine pass. H3's soundtrack comes out of the same sampling pass as the picture, and it's already decent. But if you want to re-sample the audio alone over a low denoise - cleaning up a hiss, tightening a dialogue line, re-deriving a lip-sync pass - that's this block. refine_enabled (off by default) gates it. refine_steps is 4 by default because the refine typically runs as a short low-denoise pass, not a fresh generation. refine_denoise at 0.5 is the "re-sample it alongside the picture" territory that H3 users end up in when they need the mouth to move with new audio; dropping toward 0.1-0.2 barely touches what came out. refine_sampler and refine_scheduler are the stock ComfyUI lists with euler and simple preselected, which matches what H3 wants - Turbo's contract in this same pack is Euler/simple/denoise=1.0, so staying on it keeps you in familiar territory.

The seam. seam_enabled (on) and seam_ms (80) feed the same waveform-blend logic that H3AudioSeam applies when it stitches adjacent segments: a short crossfade that removes the click at a cut and matches the level on both sides. 80ms is a sane default. Push to 150-200 when the ambience changes across the join, drop to 20-40 when a plosive or a hard consonant sits right on the cut.

The optional model socket exists for ordering. Wiring it tells the Director's scheduler that this bundle depends on the model having loaded, which is the same trick the memory-barrier nodes in this pack use.

Outputs

One: 音频设置, type MYANG_H3_AUDIO. Drop it on the Director's audio settings input. If you're hand-wiring a graph instead of using the Director, you won't get far with this node - the downstream nodes that understand MYANG_H3_AUDIO are the Director and the long-video expansion, not the stock H3 nodes H3AudioSeam plugs into.

Install

Ships in the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/civilcoco/ComfyUI-MiniMaxH3-Myang

Restart, hard-refresh, and it's under 沐阳 H3/导演台. ComfyUI Manager users: search "ComfyUI MiniMax H3 Myang". No extra Python packages. You still supply H3 weights, the Qwen text encoder, and both VAEs yourself.

The honest caveats

refine_sampler lists 44 samplers because it just forwards comfy.samplers.KSampler.SAMPLERS. That is not a licence to pick any of them - H3 was trained against a narrow contract, and the pack's own Turbo node insists on Euler with the simple scheduler. If a refine pass looks like it melted, drop back to euler/simple before you go hunting elsewhere.

Second: the refine pass works on the audio latent. If it's on and your result sounds subtly unfamiliar, that's a real re-render, not a level change. Turn refine_enabled off and compare against a plain two-segment baseline - the pack's README makes exactly this point about every speed/quality switch in it, and it applies here too.

If you're not using the Director at all and just want the picture/sound seam fixed on a hand-built chain, skip this node and wire H3AudioSeam directly. It does the same blend with visible widgets.