MiniMax H3 Semantic Anchor Bundle

Semantic anchors are Qwen-only pictures - story beats, replacement-character shots - that never consume H3's native picture/video/audio slots. The catch is that all those small anchor nodes produce a long chain of drafts, and you need one clean socket to hand the whole thing downstream. The Bundle is that socket: it finalizes any number of Semantic Picture Anchor drafts behind one graph input, owns the shared presentation size and mode, and emits a single resume-safe fingerprint.

What it does

Connect the final anchors draft from your Semantic Picture Anchor chain. Two shared settings apply to all of them:

• semantic_anchor_size - the central Qwen presentation resolution (384, 512, 768, 1024, 1280, or source). Scaling happens only when an anchor is actually called by the current prompt, so unmentioned anchors cost nothing.
• semantic_anchor_mode - timestamped_video presents #tags as timestamped Qwen Video checkpoints; picture_storyboard uses approximate Picture cues. Neither mode creates a native VAE reference - that's the whole point of the semantic path.

Optionally, connect the native references registry (your final Tagged reference chain) and the Bundle combines native @references plus Qwen-only #anchor metadata into one carrier.

Outputs:

• references - one carrier containing both native and semantic references. Connect it to Tagged Ref2VA, Plan Studio, and Run Manager.
• reference_fingerprint - the combined native + semantic incremental fingerprint for H3 Chain Plan.
• status - counts, mode, size, and fingerprint.

Where it fits

The pack's recommendation is explicit: for new workflows, register stills with Semantic Picture Anchor, chain the small nodes, and finish them with one Bundle. That one references line is then all your reference wiring needs - it's the same socket whether you have three semantic anchors or thirty. Semantic images stay outside native H3 capacity, so an inactive anchor can hang around in your graph without stealing a Picture slot.

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 - mind the MiniMax H3 Community License territory restriction.

The subtlety: presentation mode changes how H3 sees the anchors, not what they mean. timestamped_video suits anchors that are moments in a performance; picture_storyboard suits look/identity cues. If your character drifts mid-shot despite a semantic anchor, the fix is usually a #tag[timestamp] placed inside the current scene's window - an anchor timestamped outside the scene is simply ignored.