MiniMax H3 Lazy Motion AV Loader

Every frame of a motion-reference video you decode is VRAM and RAM you didn't need - H3 only ever consumes a small 24 fps scene window at a time, and the rest of the file just sits there. This loader is the lazy route: it opens the container, keeps the video file-backed (never reading the full frame tensor), and only decodes the embedded audio track so legacy AUDIO wiring has something to pull. It's a 0.4-era adapter, and the pack would rather you use Source Timeline for new work, but if you're running a saved 0.4 workflow it's exactly what you need.

What it does

Give it an absolute video_path to your motion-reference video. It produces:

• source_video - a native file-backed VIDEO wrapper. The full picture stream stays on disk; consumers pull only the window they need.
• source_audio - the complete post-skip embedded soundtrack, aligned to the same 24 fps timeline, for legacy AUDIO fan-out.
• skip_first_frames - a validated native-frame origin to connect to the tagged motion node's matching input, so the Plan timeline starts where you want without the loader actually decoding those skipped frames.
• status - resolved path, skipped origin, audio duration, and a lazy-load summary.

The skip_first_frames input has a subtlety worth knowing: it moves the timeline origin, but the VIDEO wrapper stays untrimmed. You connect the output INT to the tagged motion reference's matching input rather than expecting the video to come out shorter.

Why "legacy" and should you care

This is explicitly a Legacy 0.4 adapter. New workflows should register their motion reference container with Source Timeline, which is the 0.5 source-media contract and does the same laziness properly - recording video, audio, timing, fingerprints, and recovery data behind one typed wire. The loader exists so old graphs keep running. If you're starting fresh, reach for Source Timeline; if you're resurrecting a 0.4 workflow or just want a single container wired into a legacy graph, this is the path.

Install and gotchas

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

Restart ComfyUI, or install via Manager under "MiniMax H3 Contex Loop". No pip step - it relies on ComfyUI's own PyAV. You'll want ffmpeg on PATH as a general good practice, and a current ComfyUI with native Add Guide support (PR #15439).

The main trap: video_path is an absolute path, so a workflow shared between machines breaks until you re-point it. And because it's file-backed, don't expect IMAGE output from this node - the video stays lazy by design; you'd decode it with a preview node when you actually want to look.