MiniMax H3 Current Tagged Ref2VA Scene

Say you're building a multi-scene MiniMax H3 video where scene 3 should show the courier from scene 1's opening image, and scene 5 should reuse a character you tagged in the project. That's "prompt-driven references": you stick a @tag in the scene's prompt, and the reference that matches that tag gets pulled into the conditioning. Historically this pack made you wire it as two separate nodes - Current Shot feeding Tagged Ref2VA - plus a dozen loose output wires. MiniMax H3 Current Tagged Ref2VA Scene is the 0.6 replacement that collapses that whole pair into one node, and it's what the current reference workflows are built around.

It lives inside the scene loop. Where the loop is currently pointing, this node resolves the scene from the chain state, hands the scene's prompt and geometry to the tagged-reference machinery, and returns the conditioning and latent you render that scene with.

How it works

This is a GraphBuilder expansion node. At queue time ComfyUI expands it into three real nodes: MiniMax H3 Context Loop Current Shot (resolves the current scene), MiniMax H3 Tagged Ref2VA (does the reference-to-video conditioning), and the internal MiniMax H3 Current Tagged Scene Pack (bundles all the metadata). You never see those children on your canvas - that's the point. The pack keeps its proven internals but hides the plumbing, and gathers what used to be secondary output wires into one typed scene_data bundle. It needs the state from H3 Context Loop Start; legacy 0.4 states must keep using the old two-node route, so don't try to feed those here.

Two tag families matter. @tag is a native Ref2VA reference - a real picture that counts toward H3's active reference slots. #tag is a Qwen semantic anchor, which reinforces a character or object without counting against those slots. Both are resolved from the compiled scene prompt.

Inputs

• state - current loop state from H3 Context Loop Start.
• clip, vae, audio_vae - the H3 text encoder, video VAE, and audio VAE the tagged pass uses.
• references - the final tagged reference line, including any Semantic Anchor Bundle.
• options (optional) - a settings bundle from MiniMax H3 Tagged Scene Options. Leave it disconnected and you get native Ref2VA defaults.

Outputs

state carries the resolved scene onward, positive (CONDITIONING) and latent feed the chain's sampler inputs, and scene_data carries the typed metadata - scene prompt, seed, compiled prompt, active references, options - for the handful of cases where another node needs one of those values. The pack's own advice: don't unpack scene_data unless you actually need the value; a wire for every scene fact is how graphs get unreadable.

How to install

ComfyUI Manager (search "MiniMax H3 Context Loop") or:

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. The MultiRef companion repo is required, and because this node expands through GraphBuilder you want a current ComfyUI. Your H3 model, text encoder, and both VAEs are yours to supply.

Common issues

If you see a complaint that the node needs ComfyUI GraphBuilder, your ComfyUI is too old - update. If you get a state-type error, you're almost certainly feeding it a legacy 0.4 state; that's exactly the case where the separate Current Shot + Tagged Ref2VA nodes are still the right call. And if references silently don't apply, check your @tag/#tag names against what's in the reference nodes - under strict policy an unresolved tag is an error, not a silent skip.