MiniMax H3 Scheduled Picture Ref

H3 numbers its reference pictures positionally - <Picture 1>, <Picture 2> - and in a multi-scene loop you'll often want different faces active in different scenes. If you've been hand-editing prompts to keep those numbers straight, this node is the relief: you register a picture once under a stable @tag, pick which scenes it's active in, and the scheduler renumbers <Picture N> per scene. Remove an earlier picture and @hero_face still resolves correctly, because the numbering is recomputed from whatever is actually active.

What it does

Connect one image (Ref2VA uses only the first image of a batch), set a tag like hero_face (it becomes @hero_face in prompts - not a native number), and set scenes - blank for all, or 1, 1:4, 1,3,5:8. The previous input chains it into a schedule in your stable priority order; native numbers are only assigned after inactive entries are dropped for the current scene.

Outputs are the schedule (chain it into another entry or into Scheduled Ref2VA), a schedule_fingerprint, and a status summary.

The renumbering example the pack itself gives is the one to remember: if @picture_1 is removed or goes inactive, @picture_2 automatically compiles to <Picture 1> in that scene. That's why tags are safer than hardcoded native labels in any changing schedule.

The node never inserts prompt text - you write @picture_2 in the Plan prompt, and the scheduler only resolves the alias to the scene-local native label. The meaning of each reference (who the face is, what the look should be) still belongs in your prompt, not in the scheduler.

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 - check the MiniMax H3 Community License territory restriction.

The trap is thinking of the tag suffix as a fixed number. It's an alias, and the whole feature is that it isn't fixed. If a scene fails with an unresolved tag, check whether the picture is active in that scene and whether your selector actually covers it - a 1,3,5:8 selector means scenes 2 and 4 don't see it at all, by design.