MiniMax H3 Latent Tail Guide

Continuation video has a classic problem: how do you tell the model where the last clip ended? The usual answer is to decode a frame, shove it through conditioning, and hope the new chunk matches. This node does something better - it hands the model the raw latent tail of your previous clip, video and audio together, and calls it a day.

MiniMaxH3LatentTailGuide is the middle node in ttulttul's ComfyUI-Minimax-H3-Continuation pack, and it's the clever one. Because MiniMax H3 treats video and audio as one joint latent, the "tail" it copies is both streams at once. That's why your continuation keeps its audio in sync without a separate audio-matching pass - the sound is part of the context, not bolted on later.

How it works

ComfyUI commit e01fb4c removed H3's old first-frame/last-frame-only keyframe restriction and added native guides at arbitrary positions. This node exploits that: it slices the tail of your completed latent's video stream and audio stream, appends them to the conditioning's minimax_keyframes list at frame index 0, and hands back modified conditioning. The tail lands on the hidden overlap of your fresh continuation target, anchoring the new window to the old clip.

The "no VAE round trip" part matters more than it sounds. Decode-then-reencode loses whatever lives in the latent that doesn't survive pixels - the guide uses the exact token trajectory the previous clip ended on, which is the best possible description of where you need the continuation to start.

Inputs and outputs

• positive - the positive conditioning for the whole continuation window, coming out of your extension MiniMax H3 Image to Video node.
• previous_av_latent - the completed clip whose tail provides the context.
• target_av_latent - the fresh continuation target, straight from MiniMax H3 Guided Continuation Window.
• overlap_video_tokens and overlap_audio_tokens - the token counts describing that hidden overlap, also from the window node. They tell the guide how much of the previous latent to slice off.

The single positive output goes into your continuation BasicGuider, and from there into the sampler.

The gotcha that will waste an hour

This node only works if it sits on the active conditioning edge. The README is blunt about it: if you run the guide somewhere else in the graph but leave BasicGuider connected straight to Image to Video, the guide computes its output and that output gets discarded. You end up sampling an empty target with text-only conditioning, and the result is an unrelated, garbage continuation.

So the wiring is mandatory: Image to Video → Latent Tail Guide → BasicGuider. Not "execute it elsewhere". Also keep first_frame and last_frame disconnected on the extension node - the guide will error out if an existing guide lands inside the overlap it's claiming.

Install and requirements

Same as the rest of the pack: ComfyUI Manager (search "MiniMax H3 Continuation") or

cd ComfyUI/custom_nodes
git clone https://github.com/ttulttul/ComfyUI-Minimax-H3-Continuation

then restart. Zero Python dependencies, no model files in the pack - you just need the MiniMax H3 weights loaded normally and ComfyUI on commit e01fb4c or newer (the node checks the upstream PackedLayout signature and refuses to run on older builds).

It's honest about its limits, too: the tail guide supplies synchronized motion and audio context, but it can't guarantee a seamless semantic transition. If your scene makes a hard narrative turn, no amount of latent tail will smooth that. For straight continuation - camera keeps moving, audio keeps playing - it's the best native option H3 has right now.