MiniMax H3 Tagged Video Ref

H3's video-to-video motion transfer was the community's first "like Kling" moment when the model dropped, and MiniMax H3 Tagged Video Ref is the clean way to feed that in a multi-scene Contex Loop. It registers one reference video - plus an optional synchronized soundtrack - under a stable @tag. When any scene prompt mentions the tag (or its paired audio tag), Tagged Ref2VA activates the whole block as a native H3 <Video N> reference for that scene. Mention it in two scenes and it's used in both. Mention it in none and it sits on disk doing nothing.

This is the "keep the look and the movement" node. The video stays a native video reference, so H3 sees the full clip - appearance, action, camera - not an extracted pose. If you want to keep the performance but not the look, that's the separate Tagged Motion Ref node (which compiles the tag to a reusable <Subject N> instead). Choosing between them is the core of how this pack drives continuity.

The inputs that matter

• video - reference frames as an IMAGE batch, and it must be 24 fps. H3 is fixed at 24 fps; if your source is another rate, run it through the pack's Reference Video Prep first. The node rejects batches under 5 frames.
• tag - default performance. The prompt-facing alias.
• audio_tag - alias for the paired soundtrack. Leave it blank and it derives @<video_tag>_audio, so @performance and @performance_audio both flip the block on.
• timeline_mode - restart_each_scene (default) starts the reference at frame 0 every time the tag is used; sequential advances through the clip from the first scene that activates it. sequential needs the Current Shot state wired into Tagged Ref2VA, because the compiler needs exact scene timing.
• audio - optional, the soundtrack aligned to the same clip.
• previous - chain another reference before this one, same as the whole family.

Outputs are the family trio: references (chain it into the next reference or into Tagged Ref2VA), reference_fingerprint (the append-aware registry hash that keeps checkpoints honest about what produced a take), and status (what got registered and which timeline mode it's in).

When to reach for it

The classic use is clip continuation: you've got existing footage, you register it under a tag, and each scene prompt either calls it or doesn't. That's the "Masked AV Extension" pattern in the pack's example workflows - the accepted predecessor scene flows in as the next scene's reference. sequential mode matters there: for masked AV continuation the motion advances on delivered frames, so the repeated prefix isn't replayed after the mask boundary. If that last sentence made your eyes glaze over, start with restart_each_scene and upgrade later.

Install and gotchas

This ships with the whole Contex Loop pack - there's no standalone install:

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

Restart ComfyUI, or grab it via Manager by searching "MiniMax H3 Contex Loop" (the spelling is deliberate). Models are not bundled: you supply the H3 weights and both VAEs, plus a current ComfyUI build with native Add Guide for MiniMax H3 (ComfyUI PR #15439). ffmpeg on PATH is preferred for review and assembly, with bundled PyAV as fallback.

The gotcha that bites most people is frame rate - a 30 fps or 12 fps clip registered directly will mis-time every scene - and the H3 license geography. The MiniMax H3 Community License excludes the US, EU, UK, and South Korea from running the local weights, outputs included. The pack itself is GPL-3.0. Sort the legal bit out before you invest a weekend in a multi-scene run.