MiniMax H3 Semantic Anchors (Internal)

This is a behind-the-scenes node, and you should know that before you go looking for it in your graph. MiniMaxH3SemanticAnchorConditioning is an internal stage of the pack's Tagged Ref2VA machinery: it re-encodes the complete Qwen media presentation with your timestamped semantic Video checkpoints (or Qwen-only Picture storyboard cues) while preserving the native Ref2VA VAE/audio payload untouched. When you use the semantic anchor workflow, this node is doing the actual conditioning surgery - you just don't normally wire it yourself.

What it does

It takes four inputs:

• positive - the native Ref2VA conditioning whose media payload is preserved while the Qwen presentation tokens are rebuilt.
• clip - the H3 text encoder used to rebuild the complete multimodal presentation.
• prompt - the compiled current-scene prompt paired with the semantic presentation.
• presentation - the internal timestamped-video or picture-storyboard presentation emitted by Tagged Ref2VA's expansion.

The output is a positive conditioning with the rebuilt Qwen presentation and the original native Ref2VA VAE/audio payload, plus a status string summarizing mode, selected checkpoints, and presentation details.

The reason it exists: H3's Qwen text encoder understands a multimodal presentation - media plus timestamps - and the semantic anchor path needs to rebuild that presentation for the current scene without disturbing the native references that are already VAE-encoded. It's the boundary where "semantic" and "native" conditioning get merged back into one stream.

Should you use it directly?

Almost certainly not. The presentation input is an internal type produced by Tagged Ref2VA's expansion - it's not something you can hand-craft, and connecting it yourself means you're reimplementing the scheduler's pipeline. If you want semantic anchors in your workflow, the supported path is: Semantic Picture Anchor nodes → Semantic Anchor Bundle → the Bundle's references output into Tagged Ref2VA. This node runs inside that flow.

You might legitimately see this node referenced in error traces or while reading the pack's code; knowing what it does helps you understand why a semantic-anchor render behaves the way it does (presentation is rebuilt, native payload is not).

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 only real "gotcha" is a conceptual one: if you wire this node into a graph by hand, you've taken on the scheduler's job, and you're now responsible for the presentation being correct for the current scene. The pack's own docs mark it "(Internal)" - treat that label as a warning, not a suggestion.