MiniMax H3 Tagged Scene Options

Tagged references are how a multi-scene MiniMax H3 workflow keeps one character, prop, or location consistent: you name a picture with a tag in the project, drop @name or #name into a scene's prompt, and the right reference gets pulled in for that scene. If you're happy with the default behavior you never need this node - MiniMax H3 Current Tagged Ref2VA Scene runs fine with no options attached. MiniMax H3 Tagged Scene Options is the control panel for when the defaults aren't quite what you want, and it's designed to be shared across every scene rather than retuned per scene.

It produces one versioned options bundle that you connect to Current Tagged Ref2VA Scene's optional options socket. Because it's a typed carrier rather than a pile of widgets on the scene node, you can build it once, feed several scenes, and know they all used the same reference behavior.

The inputs that actually matter

• reference_policy (strict / soft / disabled) - how strictly @tags are validated. strict errors on an unresolved tag; soft keeps unresolved tags in the prompt with a warning; disabled passes everything through and stops enforcing. Start strict, drop to soft the first time a tag you didn't register trips you up.
• ref_image_size (match / max) - native Ref2VA reference-image sizing: size references to match the scene canvas, or cap them at a maximum geometry.
• semantic_anchor_size and semantic_anchor_mode - these tune the Qwen semantic anchors (the #tag family), not native references. Size is the presentation resolution for bare #tag and #tag[timestamp] visuals; mode decides whether a timed anchor becomes a video checkpoint (timestamped_video) or a picture-storyboard cue. Bare untimed #tags stay pictures either way.
• cache_for_upscale (default on) - save the native Ref2VA reference data so a later target-resolution pass can rebuild references at the upscaled canvas. If you ever plan to upscale a scene, leave this on; the pixel and latent upscale conditioning nodes depend on that cache existing.
• conditioning_backend (native_ref2va / external_refmod) - run references through native Ref2VA, or emit the active visuals for the external MiniMaxH3Mod RefMod path instead. You only touch this if you've installed that separate pack.
• align_audio_reference - experimental; aligns only the current tagged source-audio reference to H3's target grid. Off unless you're chasing a specific audio-boundary artifact.

Outputs and wiring

One output, options, straight into Current Tagged Ref2VA Scene's options input. Leave it disconnected for native defaults. When connected, the options get validated (bad combinations throw before they reach the sampler) and are embedded into the scene's scene_data, so MiniMax H3 Scene Data Extract can read back which options a given scene actually used - handy when you're comparing takes.

How to install

Same pack, no separate install:

cd ComfyUI/custom_nodes
git clone https://github.com/seitanism/ComfyUI-H3-Motion-Context-MultiRef.git
git clone https://github.com/ethanfel/ComfyUI-MiniMaxH3-Context-Loop.git

Restart ComfyUI (Manager: search "MiniMax H3 Context Loop"). You supply the H3 model, text encoder, and VAEs as usual.

Common issues

Two things get people. First, if #tag semantic anchors look soft or upscaled oddly, you've likely changed semantic_anchor_size when you meant ref_image_size - the two are separate knobs and only semantic visuals respect the anchor size. Second, switching conditioning_backend to external_refmod without the MiniMaxH3Mod pack installed produces an error rather than a fallback - that's intentional, since a half-wired RefMod path would silently render without your references.