MiniMax H3 Audio Reference (Plan v2)

H3's big differentiator is that it generates stereo audio jointly with the picture - no separate audio pass bolted on. That also means its prompt format is picky about what an audio reference is for. A clip of someone talking can mean "copy this voice," "copy this speech content," or "match this room tone," and those map to different H3 constructs. Audio Reference (Plan v2) forces you to pick exactly one relationship per clip, and never lets "voice" and "music style" collapse into one generic option - which is precisely the ambiguity that makes hand-written H3 audio prompts fail.

The relationships (pick exactly one)

The audio_use enum is the whole node. In rough groups:

• Voice - Voice timbre and delivery, and Dialogue or lyric content (that one needs a transcript and language because it's copying exact words).
• Music - Background-music style, Beat or rhythm.
• Effects - Sound-effect texture.
• Signal copying - Audio continuity, Copy complete signal, Copy selected part or layers, and Broad audio inspiration.

The browser UI hides controls you don't need for the selected relationship, which is genuinely nice: choose Voice timbre and the transcript field disappears.

Inputs that matter

• audio - one 2–15 second ComfyUI AUDIO segment. All Audio References in one chain share a 15-second cumulative limit. (A complete-copy/continuity soundtrack paired to a matching 362-frame video may cover that video's 15.083s padded boundary.)
• target_speaker - required only for voice timbre or dialogue/lyric content: the human alias of the Subject whose voice this is.
• language + transcript - only for exact dialogue/lyric content. The tooltip has a sharp warning: voice timbre alone must leave transcript blank so the source's words aren't accidentally reused.
• target_layer_or_event - required for music, beat, effect, continuity, and broad inspiration.
• shot_scope - optional numeric scope. For a sound effect, you also insert its <Audio N> tag inside the exact Shot sentence where the sound occurs - scope validates placement but doesn't write it.
• paired_video - connect a Video Reference handle only when this is that video's paired soundtrack; otherwise the clip routes as standalone ref_audio_N. The ordering rule: paired soundtrack immediately before its Video, standalone audio after.

Outputs

• h3_plan - continue the setup chain.
• h3_audio - the exact segment for the native route Prompt Merge assigns.
• reference_preview - role, duration, target, retention, and paired/standalone route.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/ethanfel/ComfyUI-MiniMax-H3-Guide

Restart and browse MiniMax H3 → Plan v2. No dependencies beyond a native-H3 ComfyUI build; this node is pure metadata plus passthrough.

Gotchas

• Timing the tag placement. An untagged texture/style reference falls back to the global soundscape; a tagged one lands at the exact sentence in the Shot. The pack's Foley docs hammer this: put <Audio N> in the Shot sentence where the sound is heard, or the texture will arrive at the wrong time.
• The cumulative 15-second audio limit. Stacking three 8-second clips gets rejected. Trim with the Reference Sheet nodes' non-destructive window rather than fighting it.
• Foley mode rejects copy relationships. If your plan is video-to-audio Foley (regenerating the whole soundtrack), complete-copy and partial-copy audio references are disallowed - the target stream is fully regenerated, so there's nothing to copy.

Where people get burned: they connect an audio clip, leave audio_use on the default "Choose an audio relationship," and the node refuses to compile. That's intentional - an unassigned relationship stops execution rather than guessing. Pick the role, and everything else mostly takes care of itself.