Long clips with context windows

H3 samples a clip in one pass, which limits how long a single clip can be. ComfyUI's MiniMax H3 Context Windows (ComfyUI pull request 17425, merged 2026-10-30) samples a long audio and video latent in overlapping windows, keeps audio in step with video, crops first and last frame guides to each window and applies references per window. H3 Continuity builds three things on top of it.

Status

These features are implemented and unit-tested on CPU, and were checked on one GPU setup (below). They need a ComfyUI version that includes the core Context Windows node; otherwise the nodes say so. Update ComfyUI before using any of this.

The H3 Context Windows node

Place H3 Context Windows between the model and the sampler.

Input	Meaning
context_length	Frames sampled together in one window, snapped to 17k + 5. Default 124.
context_overlap	Frames shared by neighbouring windows, snapped to 17k + 5 and kept shorter than the window. Default 39, the Timeline's default carried context.
fuse_method	How overlapping windows are blended (pyramid by default).
freenoise	Shuffle noise per window so neighbours blend better.
split_conds_to_windows	Split conditioning across windows.

Raising the clip limit

The default maximum is 362 generated frames. A Timeline may set "max_frames" (a 17k + 5 value from 362 to 991) in its JSON to allow longer clips: a clip of up to that many generated frames when fresh, minus the carried frames when continuing. A clip over 362 frames gets a note telling you to connect H3 Context Windows, because without windows the sampler would try all of it at once.

Studio limit

The Studio's editor and the Story planner still stop at 362. Set max_frames in the Timeline JSON or through the API. limits.max_frames in the prepare bundle reports the range.

Removing seams

long_merge.merge_continuing(timeline_json, first_id, last_id) turns a run of continuing clips into one clip: frames add up (still 17k + 5), prompts join in order, and previews and plate changes move by the frames before them. Cuts, Exact-length clips and a continuing clip with a first frame are refused. Windowed sampling then blends the old seams. Nothing calls it automatically.

Refining a rendered chain

H3 Join Chain reads saved segments first_segment..last_segment of one continuing chain from output/h3_sequences/<name> and joins them into a single latent, keeping each carried stretch once (0 joins through the last saved segment). Noise it with a sampler denoise below 1 and sample it through H3 Context Windows. A chain that crosses a cut, or a segment that dropped frames, is refused. The refined latent is not yet saved back as segments.

What was measured

One RTX 5090 Laptop run at 640 × 384 with an FL2VA int8 model and an 8-step turbo LoRA, windows of 124 frames overlapping 39:

Run	Result
243-frame clip with first and last guides, windowed	130 s; guides landed; no hard cut (largest frame-to-frame change 5.2, against 2.8 without windows). Slightly busier motion than unwindowed.
Same, guides unrelated to the prompt	The first and last windows follow their guide; the middle windows follow only the prompt, so the video cuts to the prompt's scene and back. Guides anchor only the window they fall in.
498-frame clip, max_frames 498, first and last guides	205 s, 20.75 s of video with audio; no hard cut; both guides land.
@image1 reference clip, 243 frames, windowed	The same woman and outfit in every window; no hard cut.
Two continuing clips merged to 226 frames, Join Chain, denoise 0.45	105 s; the join spike was gone; motion slightly busier.

A sweep of window settings (overlap, window length, freenoise, fuse_method) found no setting that was clearly better, so the defaults stay. Every windowed run was busier than the unwindowed control, so the extra motion comes from windowing itself.

Not verified: memory blocks with windows, a last-frame guide on a clip that drops frames, clips past 498 frames, higher resolutions and VRAM peaks. These results were judged from frame statistics and contact sheets; nobody has watched the clips, so judge the motion yourself.