ComfyUI-MiniMaxH3-Context-Loop/ComfyUI-MAINodes/docs/WINDOW_MODE.md

Window mode: retiming an excerpt of a longer clip

Most of this pack retimes a whole clip. Window mode is the other shape: a long clip passes through untouched except for one excerpt, which is expanded on the model's time axis (H3 Temporal Insert, or a pixel-space H3 Time Smear) and regenerated. It is the cheapest way to fix one burst in a long shot, and it is the easiest graph in the pack to edit wrongly.

The window is not one number. It lives in FOUR coupled places, and all four move together or none of them do.

The four coupled parameters

1. Prefix crop (an ImageFromBatch before the processed excerpt). length is how many frames pass through untouched ahead of the window.
2. Excerpt crop (a second ImageFromBatch). batch_index is the prefix length, length is the excerpt in world frames. The excerpt must be a legal H3 length, 17k+5, because it goes through the VAE.
3. Hold map (H3 Temporal Insert, or whatever writes the map). holds carries one entry PER WORLD FRAME of the excerpt, not per token, and the value is the local dilation. world_len is the excerpt length, and sum(holds) is the dilated frame count.
4. Conditioning length (the MiniMaxH3ImageToVideo family node feeding the regeneration). length here is sum(holds), the DILATED total, and it must also be 17k+5.

Number 4 is the one that gets forgotten, and it fails in a way that does not point at itself: a latent longer than its conditioning renders garbage, and every downstream mask is nearest-neighbour resampled to the old proportions, so the visible artifact sits at the OLD window boundary. If a window edit produces a broken render whose damage is parked where the window used to be, this is the parameter to check first.

Why "start at second X" quantizes

Write the excerpt as a head h at hold 1, a window w at hold d, and a tail t at hold 1:

excerpt = h + w + t      must be 17k+5
dilated = h + d*w + t    must be 17k+5

Subtract them and (d-1)*w = 0 mod 17. At d = 2 the window length w has to be a multiple of 17 world frames, and then h is a multiple of 17 too. So window starts land on a 17-frame lattice, which at 24 fps is 0.708 s per rung. A request to start at second 1 lands on 0.71 s or on 1.42 s; there is no rung between them. Pick one and know which way you rounded.

Procedure

1. Write the target geometry down as (prefix, h, w, t, d) and check BOTH sums against 17k+5 before you touch the graph.
2. Edit all four literals in one pass. Assert the old value as you replace each one; that is what catches a graph that has drifted from the geometry you think it has.
3. If an endpoint is fixed ("it has to end on the same frame"), solve for the start rather than sliding a fixed-length window: the lattice moves both ends.
4. Prefer window starts on a 17-multiple (frames 0, 17, 34, ...). H3 Temporal Insert recovers inserted singleton tokens that land on those anchors exactly; off-anchor ones are its worst case.

H3 Window Plan and H3 Window Collect solve a different problem (splitting one clip's dilated pass into budget-sized windows for a small card) and do their own arithmetic. This page is about hand-built excerpt graphs, where the four numbers are yours to keep consistent.