MiniMax H3 Tagged Ref2VA

This is the node the whole Tagged family feeds. Every picture, video, motion, and audio reference you registered under a @tag chains into MiniMax H3 Tagged Ref2VA, and here's where it actually does something: for the current scene, it activates only the tags that appear in the resolved prompt, renumbers those assets into compact native H3 labels (<Picture 1>, <Video N>, <Subject N>), leaves any unrelated @syntax alone, and hands the result to stock MiniMaxH3ReferenceToVideo. One shared sampling body, zero numeric reference schedules, and each scene's prompt is the entire control surface.

Think of it as the compiler step. The reference nodes declare assets; this node turns a prompt into a per-scene wiring diagram and produces the conditioning that actually drives H3.

The inputs that matter

Wire these once and they mostly stay put:

• clip, vae, audio_vae - the H3 text encoder, video VAE, and audio VAE. Yes, audio has its own VAE; stock Ref2VA needs all three.
• references - the final chain from your Tagged Picture/Video/Audio (or Semantic Anchor Bundle) nodes.
• clip_index / clip_count - connect these to Current Shot's outputs so the active refs change on each recursive iteration and schedule bounds get validated.
• prompt - the resolved current-scene prompt. Mention @hero_face and it activates; leave it out and it doesn't. This is the field that makes the whole system prompt-driven.
• width (960), height (544), length (124, H3-valid raw frame count, steps of 17 from Current Shot), ref_image_size (match limits pictures to generation pixel area; max uses H3's high-fidelity 2048px-short-edge path).

Two optional inputs are easy to miss and matter: state (Current Shot state - required if any tagged video uses sequential mode or any tagged audio uses source_timeline, because it supplies exact scene timing), and reference_policy (strict default validates sources and H3's reference capacity; soft keeps the structural checks; disabled makes pack validation warning-only and passes @tags through unchanged). Leave it on strict until something genuinely fights you.

The semantic-anchor path and upscale caching

There's a second, VAE-free route hiding in the optional inputs. Write a tagged picture as #hero_face[2.50s] and, depending on semantic_anchor_mode, it becomes either a timestamped Qwen semantic checkpoint (timestamped_video) or a Qwen-only <Picture N> storyboard cue whose timestamp is written as an approximate prompt instruction (picture_storyboard). Neither creates a VAE reference, and neither is a hard frame or spatial lock - it's semantic guidance. semantic_anchor_size (512 default) trades Qwen token cost against visual detail.

And cache_for_upscale (default on) automatically saves the native reference latents, compact Qwen presentation frames, and original picture masters keyed to the checkpoint fingerprint - that's what lets the pack's deferred upscale loop rebuild pass-2 references without the original reference wires.

Outputs: positive (CONDITIONING into the H3 sampling body), latent (the empty H3 AV latent), compiled_prompt (the exact prompt H3 sees after tags compile - great for debugging), active_references (the scene-local native mapping), and reference_fingerprint.

Install

Part of the Contex Loop pack:

cd ComfyUI/custom_nodes
git clone https://github.com/ethanfel/ComfyUI-MiniMaxH3-Contex-Loop.git

Restart ComfyUI (or Manager → "MiniMax H3 Contex Loop"). Needs a current ComfyUI build with native Add Guide for MiniMax H3 (PR #15439) and the GraphBuilder core it depends on. No models bundled - supply the H3 weights and both VAEs. ffmpeg on PATH preferred for review/assembly; PyAV fallback. Pack is GPL-3.0; H3 weights are territory-restricted (US/EU/UK/South Korea, outputs included). If a reference silently never shows up, check the resolved prompt first - the compiler only activates what the prompt names.