MiniMax H3 Semantic Picture Anchor

Native H3 reference slots are a finite budget - nine active pictures, and every one you spend on a story beat is one you can't spend on identity. Semantic Picture Anchor is the workaround: it registers an image under a #tag that only H3's Qwen text encoder sees, as a timestamped semantic cue, without ever becoming a native VAE reference or consuming a slot. It's the cheap way to reinforce "this is the replacement character, and here's the beat where she appears" without blowing your reference budget.

What it does

Connect one image and give it a tag - like story_beat, which becomes #story_beat[2.50s] when you write it in a scene prompt (don't include the # in the tag field itself). The previous input chains it into the growing semantic-anchor draft; chain as many as you want, then connect the final draft to one Semantic Anchor Bundle.

Outputs: anchors (the unscaled draft, for another anchor or the Bundle) and status (registered tag and draft size).

The key property is in the naming: these are Qwen-only pictures. They're presented to the Qwen text encoder as approximate semantic checkpoints at a scene-local timestamp - #replacement[0.00s] #replacement[2.50s] #replacement[4.75s] - which is useful for reinforcing identity later in a shot. But they are not exact frames, not pose or spatial-mask control, and not motion. If you need those, that's the native path (@tag references), which does consume slots.

The scaling is deferred, not eager: anchors store the original picture, and only scale to the presentation size when a scene actually calls the #tag. So you can keep a big library of inactive anchors in the graph with zero cost.

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 treating #anchor[time] like a stronger control than it is. It's an approximate Qwen presentation cue, and the timestamp must fall inside the current scene - a timestamp outside the scene window is ignored. If a character drifts, the anchor reinforces, it doesn't hard-lock. Pair the # semantic anchor with a native @ reference when you need real identity fidelity and can spend the slot.