MiniMax H3 Tagged Motion Ref (Source Timeline)

The Contex Loop pack's 0.5 release reworked how source media enters the graph, and MiniMax H3 Tagged Motion Ref (Source Timeline) is the primary motion-reference node of that rewrite. Instead of taking a raw video, it takes one typed H3_SOURCE_TIMELINE wire from the pack's Source Timeline node - the thing that describes a master file once (video, audio, native FPS, skip origin) and lets every downstream consumer request just the window it needs. This node registers that timeline as a reusable @motion subject without ever decoding the full video or the full audio track. Tagged Ref2VA and the preview node decode only the active scene's window.

The pitch is simple: describe the source once, reference it everywhere, and never pay to decode the parts you don't render. On a long master clip - the kind you'd actually want for a multi-scene production - that's a workflow that stays snappy instead of choking on a 5-minute IMAGE tensor.

Inputs that matter

The motion guts are shared with the eager and path variants, but the entry point differs:

• source_timeline - one wire from MiniMax H3 Source Timeline. That node is where you point at the file, set embedded/external audio, and define the frame origin (skip_first_frames). If you haven't built one, do that first; this node refuses to invent a timeline for you.
• tag (default motion), target_subject (must be existing native labels like <Subject 1>), motion_description (transferable action evidence only - no identity, wardrobe, or setting talk) - the same motion fields as its siblings, same validation.
• reference_short_edge - scene-local decode size. Default 384 keeps coarse pose and drops appearance tokens; source preserves original resolution. Source media stays file-backed at its full resolution either way.
• paired_audio - off (default) never sends paired reference audio; embedded sends the Source Timeline audio window alongside the motion. Reference-local choice, so the final assembled track is unaffected.
• audio_tag - alias for the paired audio; blank derives @<motion_tag>_audio.
• timeline_mode - sequential (default) follows Plan timing; restart_each_scene starts at source frame 0 each activation.
• previous - chain the registry, as always.

Outputs: the standard references / reference_fingerprint / status, plus preview_source, an H3_LAZY_MOTION_SOURCE descriptor so the scene preview can show the active window without materializing the source.

Which motion node should you use?

If you've already built a Source Timeline (which 0.5's own example workflows do), this is the one - it's the primary path, and the timeline is the connective tissue between Plan, Run Manager, previews, Loop Start, and Assemble. If you're on an older 0.4-style graph or just want a direct file/VIDEO route without standing up a timeline, the Path variant (Lazy VIDEO/Path) covers you. The eager IMAGE-batch version is the compatibility-minded fallback and the least RAM-friendly of the three. For masked AV continuation, both lazy variants use delivered-frame timing so the repeated prefix isn't replayed past the mask boundary.

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"). No models bundled - supply the H3 weights and 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; bundled PyAV is the fallback. The pack is GPL-3.0, but the H3 weights under it are territory-restricted (US/EU/UK/South Korea excluded, outputs included), so read that license before you commit a production to it.