H3 Tail Context (alpha)

H3 Tail Context is the handoff node in the MAINodes segment-chaining story. A long clip built as a chain of H3 generations only works if each new segment knows how the previous one ended - the last pose, the last background state, the last few seconds of audio. That's a guide, in H3 terms, and this node extracts it: the last handle frames of a finished, recovered segment plus the sample-exact matching audio, shaped exactly for MiniMaxH3AddGuide at frame_idx 0 of the next segment.

What "the seam" means here

The pack's own framing is blunt: "Full H3 VAE on the way in, no noise, no TAE: this is the seam." No noise means it's not a v2v init - these frames are handed over as clean reference conditioning, not as a starting point for denoising. Full H3 VAE means it goes through the real encoder, not the cheap temporal-alignment extractor, because the guide needs to be in latent space the model actually reads. The result is that the continuation starts from a place, not from a guess.

Inputs and outputs

Inputs are three and two of them are boring:

• images - the finished (recovered) segment's frames.
• plan - the segment plan (the same plan object H3 Trim and the other assembly nodes read; it tells the node where the handle lives).
• audio (optional) - the segment's track; wire it and you get audio back out.

The outputs are the interesting part:

• tail_images / tail_audio - the extracted handle, ready for MiniMaxH3AddGuide.
• guide_frame_idx - the frame index to pass as the guide's anchor.
• handle - the handle spec, for the assembly/trim side.
• report - what was extracted and how it lines up.

You'll typically chain this into H3 Seam Normalize too: tail_images is exactly the source_tail calibration reference that node wants, and H3 Scene Color Stats expects a delivered tail as well. One handoff, three consumers.

Alpha honesty

This corner of the pack is new (2026-08 additions) and the README marks it alpha with the usual caveats: names and defaults may move, and playback mileage is thin. The guide path itself is the validated pattern - H3 supports image+audio guides natively, which is a big part of why the pack can anchor audio across a join at all - but the extraction math here is young. If a seam drifts after the handoff, don't assume the node is broken; check that the handle length actually matches what the continuation was given.

Install

It comes with matlowai/ComfyUI-MAINodes:

cd ComfyUI/custom_nodes
git clone https://github.com/matlowai/ComfyUI-MAINodes

Restart, find it under MAINodes/alpha. No extra Python deps; you need MiniMax-H3 weights and ComfyUI's H3 support (check the licence's territory carve-outs for the US, EU, UK and Korea before committing). The example graphs also expect ComfyUI-KJNodes for a few patches, though Tail Context itself has no dependency on it.

Where people stumble: forgetting that the handle is hidden context - it's material the continuation renders and later throws away, so you can't see it in the final output. If the seam is wrong, look at the handle, not the visible frames.