MiniMax H3 Tagged Motion Ref (Lazy VIDEO/Path)

Every other reference node in the Contex Loop pack materializes your media as an IMAGE tensor the moment it runs. MiniMax H3 Tagged Motion Ref (Lazy VIDEO/Path) refuses to. It stores a file descriptor - the path plus a fingerprint - instead of a decoded frame batch, and lets Tagged Ref2VA decode, convert to 24 fps, and resize only the exact time window of the current Plan scene, on demand. For a long reference clip, that's the difference between a workflow that sips RAM and one that drags a 5-minute video into VRAM every queue. This is the node you reach for when your motion source is big and your patience is short.

It's also the honest answer to a real trap in this pack: if you feed a raw video path to the eager Tagged Motion Ref, you pay for the whole decode up front. The lazy version exists precisely so you don't.

The two ways in, and the timing knobs

You normally don't type a path at all. Connect core Load Video (ComfyUI's native file-backed loader) to source_video - the node reads only the file path and never calls get_components, so that same loader output can fan out to Run Manager as your source video asset. The video_path string is the direct-path fallback for when you don't have a native VIDEO wire handy. Either route keeps the full frame tensor on disk.

The timing controls are where this node earns its keep:

• skip_first_frames - ignore this many frames at the source file's native FPS before the Plan timeline begins. The decoder seeks near the point; skipped frames never become tensors at all. Great for chopping a logo or a false start off the front of your reference without re-encoding it.
• timeline_mode - defaults to sequential here (unlike the eager motion node), so the reference follows Plan timing, which is the right default for masked AV continuation.
• use_embedded_audio - decode the exact same time window from the video's single embedded audio stream and pair it with the motion reference as Ref2VA audio. Turn it off if the source has music you don't want leaking into the generation.
• reference_short_edge - 384 default (minimal RAM and appearance pressure), up to source for the original resolution.

It still shares the motion machinery with its eager sibling: target_subject must be existing native labels like <Subject 1>, motion_description is transferable action evidence only, and @tag compiles to a reusable <Subject N>.

Outputs that matter

Beyond the standard references / reference_fingerprint / status, this node adds two: preview_source (an H3_LAZY_MOTION_SOURCE descriptor that lets the scene preview node show the active window without ever building a full IMAGE batch) and source_audio (the complete post-skip soundtrack, for legacy source-track wiring - empty when embedded audio is off and no AUDIO came in). If you're feeding the full track onward for chain alignment, the pack's own docs say to prefer Lazy Motion AV Loader so the native VIDEO can also fan out to Run Manager while its post-skip source_audio drives source-track chain alignment.

Install

Same pack, same dance:

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

Restart ComfyUI (or Manager → "MiniMax H3 Contex Loop" - the spelling is deliberate). No models bundled: you supply the H3 weights plus video and audio VAEs, and a current ComfyUI build with native Add Guide for MiniMax H3 (PR #15439). ffmpeg on PATH is preferred for review and assembly; bundled PyAV is the fallback. Pack is GPL-3.0; H3 weights are territory-restricted (no US/EU/UK/South Korea, outputs included) - check the license before building anything on it.