MiniMax H3 Tagged Audio Ref

MiniMax H3 generates video and its audio together - native stereo sound, not a bolted-on music pass. That's the headline capability of the model, and MiniMax H3 Tagged Audio Ref is how you feed it a reference voice, a piece of music, or a sound cue in this pack's prompt-driven system. Register an AUDIO value under a stable @tag (default voice), and the audio is sent to H3 only when the current scene's prompt contains that tag. Same mental model as the picture and video reference nodes: the prompt decides what's audible, the graph stays quiet.

Where it gets interesting is the second mode. Because audio is how H3 keeps sound connected across scenes, this node can also hold the full Loop source track and let Tagged Ref2VA derive the exact per-scene timeline slice - which is how you get a soundtrack that advances correctly through a multi-scene production without creating a fingerprint cycle or a circular Current Shot link.

The one decision: timeline_mode

• standalone (default) - sends the AUDIO value you connected unchanged whenever @tag is active. Use this for a fixed voice line or a static music cue that should sound the same every scene it appears in.
• source_timeline - treats the connected audio as the same full source track used by Loop Start. Tagged Ref2VA derives the exact current-scene slice inside itself. Use this when your audio is the production soundtrack and scenes are chunks of it.

If you pick source_timeline, there's one more toggle: align_audio_reference applies the same optional 15.070s-safe H3 audio-grid cap that Current Shot uses. H3's audio encoder runs on a fixed grid, and a reference whose duration lands just off that grid can encode a token shorter than the stock target - the alignment knob snaps it to the safe cap. Importantly, it only affects the reference used during sampling; the full source track and the final assembled audio are never modified. If the standalone-mode node ever throws "received no AUDIO value," it's because Current Shot has Source reference off and source_audio_slice came back empty - connect a loader directly.

The rest is the family standard: tag (default voice), optional previous to chain the registry, and the trio of outputs - references, reference_fingerprint (the checkpoint-safety hash of the ordered registry), and status (tag, timeline mode, alignment policy, source count).

Where it leads

Chain the references output into Tagged Ref2VA along with your picture/video references. It encodes active audio references with the H3 audio VAE (a separate VAE from the video one - the Ref2VA node takes both). Mention @voice in a scene and the voice lands in that scene's conditioning; the pack's "Ref2V Studio Tagged Source Audio" workflow shows the full source-track arrangement.

Install

It ships with the Contex Loop pack:

cd ComfyUI/custom_nodes
git clone https://github.com/ethanfel/ComfyUI-MiniMaxH3-Contex-Loop.git

Restart ComfyUI (or Manager → "MiniMax H3 Contex Loop"). No models bundled - you need the H3 weights and both VAEs, plus a current ComfyUI build with native Add Guide for MiniMax H3 (PR #15439). ffmpeg on PATH is preferred for review/assembly; PyAV is the fallback. Pack is GPL-3.0; the H3 weights are territory-restricted (no US/EU/UK/South Korea, outputs included) - worth checking before you build a soundtracked production on it. And if your audio timing ever looks half-a-frame off, the 15.070s grid is the first suspect.