H3 Extend Window

Anyone who has chained video clips by hand knows the drill: generate 81 frames, take the last frame, feed it into the next generation as an image, repeat, and watch the drift accumulate - contrast creeping up, faces slowly becoming someone else. It works, and you spend the whole project managing it. H3 Extend Window is the organised version of that loop for MiniMax H3.

It takes the clip so far and opens a window for the next segment: a latent with the clips's tail already in it, held rather than re-generated, so the sampler only has to produce what follows. Send that window to your sampler, then send the sampler's result to H3 Extend Append. Loop, and you have a long take built from short segments.

The five continuity modes

continuity is the node's whole personality, and it's the field worth reading twice.

carry is one unbroken shot. It copies the clip's last frames along with the soundtrack under them into the window and masks them, so the sampler holds them and generates only what follows. It's the default, and it's the only mode that doesn't need a VAE.

refresh does the same but softens the detail those frames gained before carrying them - the fix for a segment that comes back crunchier than the one before it. refresh_gain (default 1.15) is how much detail a pass adds, and the refresh brings the carried frames below that so the pass lands back on the opening's reading; 1.0 softens to match the opening exactly.

handoff starts the next segment on the carried frames' last frame alone - a cut that opens exactly where you were. reference hands the whole tail over as a video reference instead, which is the one for keeping a cast and a place when the scene jumps, and it wants at least 56 frames of overlap to have anything to work with.

cut, or an overlap of 0, samples a new scene from an empty latent of its own length with nothing carried at all. Clean scene change, no continuity.

refresh, handoff and reference all need vae wired; carry and cut don't.

The numbers, and the grid they live on

H3 frames sit on a 17k+5 grid - 5, 22, 39, 56 - and overlap_frames snaps down onto it, defaulting to 22. Longer overlap gives the new frames more of the scene to continue from; shorter gives the sampler more room to invent.

extension_frames (default 102, snapped down to a multiple of 17) is how many new frames this pass adds, which at 24fps is about 4.2 seconds. 17 is roughly 0.7s. You're buying these in steps of 17 whether you like it or not.

pass_index is which segment this is, from 0, and the intended wiring is a While Loop's index straight into it. Wire the same index to H3 Extend Append and the two nodes stay in step without you tracking anything.

Two optional seeds get at the same problem from different ends. drift_control takes back some of the contrast and fine detail the carried frames have gained, measured against the clip's opening frames - 0.0 carries them exactly as sampled, 1.0 takes all of it out, and it only ever softens, never sharpens. audio_release opens the held soundtrack back up over where it meets the new frames, in audio latent steps (8 is about 0.2 seconds at 40 steps/sec, 0 is a hard edge).

If you wire prompts - the bundle from MiniMax H3 Conditioning - it supplies this pass's prompt and its frame count, and extension_frames stops being read. That's the shortcut for a run whose rows you already wrote.

Outputs: window (goes to the sampler's latent input), positive (this pass's prompt, for the guider), overlap_frames (feed the same number to H3 Extend Append), and a report string stating what the pass will sample and what it snapped to. Read the report when a segment comes out the wrong length.

Installing it

Part of WAS Node Suite v3 (MIT, WASasquatch), so:

cd ComfyUI/custom_nodes
git clone https://github.com/WASasquatch/was-node-suite-comfyui.git

Or ComfyUI Manager, search WAS Node Suite v3. ComfyUI 0.14.0+, Python 3.10+, no pip packages, no bundled weights. Config, wildcard and LUT folders land in <ComfyUI user dir>/was-node-suite/ on first start, which is why that first launch is slower. The workflows folder in the repo has a working extend-loop example - docs/workflows/minimax-h3-extend-loop.json - and it's a much better starting point than wiring this blind.

The reality check

Chaining segments is still chaining segments. Overlap and drift control push the ceiling up; they don't remove it, and it's the same ceiling every chunked video tool has. Decode once, after the last segment - decoding per pass is how you end up with a stitched clip whose colour shifts at every join. And check where you stand on H3's community licence before you build on it: the open weights are geofenced out of the US, EU, UK and South Korea, and the exclusion reaches the clips you make with them.