MiniMax H3 Tagged Motion Ref

Here's the node that does the clever trick: MiniMax H3 Tagged Motion Ref registers a video as action evidence - the pose sequence, the gesture, the timing - and compiles its @tag into a distinct reusable <Subject N> whose performance is then transferred onto an existing character you name. The physical clip stays in the graph as <Video N>, but the prompt-facing tag becomes a motion role. So you can write "the dancer's routine" once, point it at your hero subject, and have scene after scene reuse that same performance on a different character, wardrobe, or set - without dragging the source video's appearance along.

Read that last bit carefully, because it's the honest caveat: this is semantic motion transfer, not a pose extractor. There's no skeleton, no keypoints. H3 watches the clip and reinterprets the movement. It's very good at it - community reaction to H3's V2V was genuinely "like Kling" - but it's not pixel-locked to the source, and you shouldn't expect frame-accurate reproduction of a stunt.

What makes the motion transfer work

The default reference_short_edge of 384 is the knob that does most of the work. It downsamples the video before it ever becomes a reference, which strips appearance pressure (source identity, wardrobe, lighting) and keeps coarse pose and temporal movement. Bump it to 512 or 768 if you need more fidelity, or source for an exact baseline. The motion_description field is the other half: describe only the transferable performance - "the supplied pose sequence, action, and motion timing" is the default - and explicitly not identity, wardrobe, setting, or composition. The node validates target_subject too: it must contain only existing native labels like <Subject 1> (it raises if you type anything else), and the description can't be blank.

The rest is the family standard: video must be a 24 fps IMAGE batch of at least 5 frames (use Reference Video Prep for other frame rates), tag defaults to motion, audio_tag derives @motion_audio if blank, and timeline_mode is restart_each_scene (frame 0 every activation) or sequential (advances from the first scene that uses the tag, which is what you want for masked AV continuation so the repeated prefix isn't replayed). Optional audio pairs a soundtrack to the motion clip, and previous chains the registry. Outputs are references, reference_fingerprint, and status like every node in this family.

Where it fits in a workflow

You chain its references output into Tagged Ref2VA, which compiles the tag into the native subject and encodes the clip with the video VAE. If your reference video is long, or you're on a RAM diet, the Path and Source Timeline variants of this same node decode only the active scene window instead of materializing the whole clip - check those before you assume this eager IMAGE version is the one to use for a 5-minute master.

Install

It's part of the Contex Loop pack, no standalone install:

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

Restart ComfyUI (or Manager → "MiniMax H3 Contex Loop"). The pack ships no models - you supply the H3 weights and its video + audio VAEs, and a current ComfyUI build with native Add Guide for MiniMax H3 (PR #15439). ffmpeg on PATH is preferred for review/assembly; PyAV is the fallback. The pack is GPL-3.0; the H3 weights beneath it are territory-restricted (no US, EU, UK, or South Korea, outputs included), so check the license before you build a pipeline on it.