MiniMax H3 Append Continuation

Every MiniMax H3 continuation samples a window that's bigger than what you actually want to keep - the leading overlap is just motion context. This node is what happens after sampling: it removes the overlap, keeps your original clip exactly as it was, and welds on only the newly generated video and audio tokens.

MiniMaxH3AppendContinuation is the third node in ttulttul's ComfyUI-Minimax-H3-Continuation pack, and it's the payoff step. The guided window plans, the latent tail guide provides context, the sampler does the work - and this node assembles the result into a single cumulative latent you can decode and save.

How it works

It takes two latents: your completed initial clip (previous_av_latent) and the freshly sampled window (sampled_window). It trims the overlap off the head of the sampled window - in the shipped workflow that's the leading 22 frames - and concatenates the remaining new tokens onto the end of your original clip, for both the video stream and the audio stream.

The part that's easy to underrate is the validation. The node checks that the sampled window matches your original latent in spatial shape, dtype, and device, and it refuses to append if the numbers are off. Because H3's video runs on a 24 fps frame grid while its audio latent runs on a 40 Hz timeline, the audio counts shift depending on where you are in the clip (a 119-frame extension can add 198 or 199 audio tokens, depending on position). The node computes those from the global frame boundaries, so appending repeatedly doesn't slowly drift your audio out of sync. That's the quiet reason this pack exists - it's the drift prevention that makes multi-segment clips viable.

Inputs and outputs

• previous_av_latent - the completed cumulative clip, same latent you fed the window node.
• sampled_window - the continuation window after it's been through KSampler or KSampler Advanced. Don't feed it the pre-sampling target; that's the whole point of the sample.
• overlap_video_tokens / overlap_audio_tokens - the overlap counts from MiniMax H3 Guided Continuation Window, telling this node how much head to cut off.
• transition_video_tokens / transition_audio_tokens - always zero in the current release; the shipped workflow discards the entire overlap.

Outputs: latent (your original clip plus the new suffix - wire this into ComfyUI's native VAEDecode and VAEDecodeAudio, then CreateVideo/SaveVideo) and total_length, an INT of the accumulated frame count, handy for tracking how long your scene has grown.

The honest limitation

The splice is hard. Because the overlap is discarded entirely, there's a clean cut between your original clip's last latent token and the first new one - and a clean cut in latent space can show up as a flash, a hitch, or a visual jump after decoding. The tail guide gives the new window motion and audio context, but it can't guarantee a seamless handoff. The author says it plainly: continuation quality still depends on the model, the prompt, the seed, and your sampling settings. If your first attempt has a visible seam at the cut, that's the expected failure mode, not a broken install.

There's also a memory reality check: the cumulative latent keeps growing with every append. Bounded window, growing scene - that's the trade, and it's fine for a few extensions but worth knowing before you chain twenty of them.

Install

Manager search "MiniMax H3 Continuation", or:

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

Restart and you're done - no pip dependencies, no extra model files beyond the H3 weights themselves, and ComfyUI commit e01fb4c or newer required. The best way to see all three nodes working together is to load the pack's own example_workflows/guided_continuation.json: a 243-frame initial clip, a 22-frame hidden overlap, 119 new frames, roughly five more seconds on the timeline. That's the workflow these nodes were built for, and it's the fastest way to confirm your wiring before you start experimenting.