MiniMax H3 Contex Loop Trim

Every H3 chain scene comes out with a pinned overlap at the head - repeated context frames that have to come off before you concatenate scenes. Trim only the images and you've just shoved the soundtrack ahead of the picture by the same span: at 5 frames that's 208ms, inaudible on ambience but squarely offbeat on anything with a pulse. This node removes the leading pinned frames from both picture and sound so the clip is actually usable, and in 0.5 chains it resolves the per-scene blend from state instead of making you wire a constant.

What it does

Feed it the decoded images from the current H3 sample and a trim_frames count - normally the trim_frames output of MiniMax H3 Contex Loop Context, which knows what the encoder actually produced. Connect audio if the clip has sound; it gets trimmed by the matching duration. The fps (default 24) converts the frame trim into an audio duration and must match what you feed Create Video.

Outputs:

• images - the delivered frames with the repeated leading context removed.
• audio - trimmed by the same duration. With match_tail on (the default), it also time-conforms small H3 audio-grid mismatches so duration equals frames/fps exactly - H3's rounded 40 Hz audio grid runs about 8ms off picture, and that error accumulates down a long chain.
• images_with_overlap + overlap_frames - an optional blend-ready stream retaining only the final part of the repeated visual context, for external stitchers that need an overlap count.

In a 0.5 chain, the recommended wiring is to connect Current Shot's state output into state. Then the node reads the active scene's resolved blend directly from the Plan, so a per-scene value can never be overridden by a Plan default. When state is connected, the retain_overlap_frames integer is ignored (it's the legacy manual path for external stitchers).

Install and gotchas

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

Restart, or install via ComfyUI Manager under "MiniMax H3 Contex Loop". No pip step; a current ComfyUI with native Add Guide (PR #15439) is expected, and ffmpeg on PATH is preferred. Models aren't bundled - mind the MiniMax H3 Community License territory restriction.

The classic mistake is trimming images only and letting the audio run long, which is exactly the desync this node prevents - so don't bypass it with a plain image slice. And if your audio drifts later in the chain, check match_tail; it's the fix for H3's ~8ms grid rounding, not a band-aid on your mux settings.