MiniMax H3 Tagged Picture Ref

If you're building a multi-scene MiniMax H3 video with this pack (the one with "Contex" misspelled on purpose, yes, that one), the whole point of the Tagged reference family is that the prompt decides what's on screen, not the graph. MiniMax H3 Tagged Picture Ref is the simplest member of that family: you hand it one reference image, give it a stable @tag, and from then on the picture only activates in a scene when that scene's prompt actually mentions the tag. No numeric reference schedule to babysit, no "picture 1 is always in" logic.

It's a registration node, not a generation node. Nothing is sampled here. It appends the image to a chain of H3_TAGGED_REFERENCES and lets the pack's compile step figure out native labels later. Think of it as pinning a face to a name the prompt can call.

What you actually set

Only a couple of fields matter for a first run:

• image - your reference picture. If you accidentally connect a batch, only the first frame is used, so don't sweat it.
• tag - the alias the prompt will use, default hero_face. Write @hero_face in a scene prompt and the picture is active for that scene; leave it out and it's ignored.
• previous - optional. This is how you chain multiple references: feed the references output of one Tagged node into previous of the next. Order matters, because it becomes the priority order for scene-local native numbering.

The outputs are the same trio across the whole family: references (the registry, chain it onward), reference_fingerprint (a hash of the ordered registry - this is how checkpoints know which references produced a take, so a later resume doesn't silently recompile with different media), and status (a human-readable summary of what registered).

Where it leads

The registry you build here terminates in MiniMax H3 Tagged Ref2VA, which rewrites @hero_face into H3's native <Picture 1> label for the active scene and encodes it with the video VAE. There's a second path worth knowing about: write the tag with a timestamp, like #hero_face[2.50s], and you get a semantic-only cue - a Qwen visual checkpoint at 2.5 seconds into the scene with no VAE reference and no hard frame lock. That's the legacy compatibility path, and the pack's own docs steer new Qwen-only images toward the Semantic Picture Anchor + Bundle nodes instead. If you just want a reference face to persist across scenes, plain @tag is the honest tool.

Installing and the fine print

This ships in ethanfel/ComfyUI-MiniMaxH3-Contex-Loop, which started life as NikoDemon80's H3 Motion Context and grew into a full checkpointed production loop. Install is the usual custom-node dance:

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

Restart ComfyUI (or use Manager and search "MiniMax H3 Contex Loop"). The pack bundles no model files - you need the MiniMax H3 weights plus its video and audio VAEs selected in the workflow, and a current ComfyUI build that includes native Add Guide for MiniMax H3 (ComfyUI PR #15439). ffmpeg on your PATH is preferred for review/assembly; the bundled PyAV is the fallback. And before you commit to a long project, remember the H3 Community License itself: the local weights and their outputs are geofenced out of the US, EU, UK, and South Korea. The pack is GPL-3.0; the model underneath has its own map.

One trap beginners hit: this node is a picture reference. If you feed it a video expecting motion transfer, nothing moves - you want the Tagged Video or Tagged Motion nodes for that. For stills, it's about as low-friction as reference conditioning gets.