MiniMax H3 Reference Sheet Audio Reference (Legacy)

H3 generates audio jointly with the picture, and it can do more with a sound reference than "sound like this." It can copy a track wholesale, borrow a voice's timbre, match a music style or beat, or follow a sound texture - and each of those is a different prompt decision. This node is the legacy bridge that takes the audio you selected in a Reference Sheet gallery, pins down which relationship you mean, and hands the decoded clip to H3's native standalone ref_audio_N input.

It's the audio half of the sheet-reference family. Small, single-purpose, and easy to wire wrong in one specific way.

What you set

• reference_sheet - connect the Reference Sheet's reference_sheet output; the gallery selection is used automatically.
• audio_use - the relationship. Four choices: reference voice/music/beat/sound (imitation, no copying), reuse the complete signal (1:1 copy), reuse only part or selected layers (copy with a new mix), or broad audio mood (weak reference).
• shot_scope - optional workflow-only scope for this audio relationship.
• notes - workflow-specific instructions. The saved sheet description stays unchanged.

previous_audio_context chains from another sheet-audio reference when you have more than one, preserving order.

Outputs

audio_context chains to the next sheet-audio reference; connect only the final one to the Prompt Guide's audio_context input. h3_audio is the decoded saved audio for the native standalone ref_audio_N socket listed in routing_report - that report is where you find which number your clip got and whether the route matches your intent. routing_report also carries labels, relationships, durations, and shot scopes.

Installing

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

Restart ComfyUI; it's under MiniMax H3/Reference Sheets. No Python dependencies, no model downloads - just a ComfyUI build with native MiniMax H3 support.

Where people get burned

Two things, both documented in the tooltips and both worth taking literally.

First, Qwen receives saved text metadata, not waveform analysis. The enhancer can't listen to your clip. If you want the generated audio to actually reflect what's in the file, the description you wrote on the sheet and the notes here are the entire channel - be specific, and say things like "calm female voice, soft reverb" rather than "the file."

Second, audio can't be the only Ref2VA media input. H3 refuses to run reference generation on audio alone, so this node needs a visual reference_context in the same workflow. People hit this, blame the node, and the node is just relaying the model's constraint.

Also remember the arithmetic: H3 allows at most three audio clips and 15 seconds of cumulative reference audio, and a trimmed segment from the sheet still counts toward that limit. When the total is tight, keep audio_use on the imitation options - a full 1:1 copy eats the entire budget by design.