MiniMax H3 Scheduled Audio Ref

Scheduling references per scene in H3 has a nasty failure mode: native labels are positional. Add a picture at the front of your chain and every <Audio 2> in your prompts silently points at the wrong track. This node is the fix - you tag your audio once with a stable alias like @voice, say which scenes it's active in, and let the scheduler renumber <Audio N> labels per scene so your prompt text never goes stale.

What it does

Connect your audio, give it a tag (a stable alias such as @voice - note it is not a native Audio number), and set scenes - blank for all scenes, or selectors like 1, 1:4, 1,3,5:8. Then chain it into the rest of your scheduled references via previous (priority order only; it never fixes a permanent native label).

Two outputs matter:

• schedule - the reference schedule to chain into another entry or into Scheduled Ref2VA.
• schedule_fingerprint - a SHA-256 of all sources, tags, and selectors, for checkpoint safety.

The renumbering is the core idea: only audio references active in the current scene consume <Audio N> numbers, and they're numbered compactly from <Audio 1> each scene. So if an earlier track is removed or goes inactive, @voice still compiles to the right native number wherever it's used. That's the whole reason tags exist.

A couple of sharp edges. This node inserts no text into your prompt - you write @voice in the Plan prompt yourself, and the scheduler only resolves the alias at compile time. And for a video's synchronized soundtrack, don't use this node: use the paired audio socket on Scheduled Video Ref instead, which keeps the track index-paired with its video.

Install and gotchas

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

Restart, or install via ComfyUI Manager under "MiniMax H3 Contex Loop". No pip dependencies; a current ComfyUI with native Add Guide (PR #15439) is expected. Models aren't bundled - see the MiniMax H3 Community License for the territory restriction.

The classic gotcha is treating the tag suffix as a fixed native number. @voice is a convenience, <Audio 1> is whatever this scene says it is. If a checkpoint verify fails after you add an earlier reference, that's the scheduler working as intended - active changes invalidate dependent scenes, and future-only additions shouldn't.