ComfyUI Node
H3 Sliding Window
Restricts each latent frame's attention to a sliding window of nearby frames plus a rotating set of long-range taps, leaving text and reference rows fully attended. Turns attention's quadratic duration scaling into near-linear: measured 1.7x at 5s, 3.2x at 10s, 6.2x at 20s. Needs a Triton-capable GPU; falls back to dense attention if flex_attention is unavailable.
H3 Sliding Window
- model
- MODEL
◄window8►
◄follow_taps2►
◄jitter1.00►
◄isolate_referencestrue►
◄start_percent0.00►
◄end_percent1.00►
CategoryNynxz/H3/Sampling
Inputs (7)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | A MiniMax H3 model. Passed through, patched. | |
| window | INT | 81–512 | Latent frames each side that stay densely attended. The main dial: cost falls roughly as the window shrinks, and local motion is what breaks first when it is too small. A window that covers the whole clip is an exact no-op. 8 is ~2s of context each way at H3's frame rate. |
| follow_taps | INT | 20–16 | Extra frames outside the window each frame attends to, placed across everything the window excludes and rotated every layer — so 50 layers cover the whole clip while any one layer pays for 2. Raise if the clip loses global consistency (drift, colour or identity wandering over time). 0 is a pure local window. |
| jitter | FLOAT | 1.000–1 | How far the follow taps rotate between layers. 1.0 sweeps the clip with a golden-ratio offset so no two layers tap the same frames; 0.0 pins every layer to the same fixed taps, which is the A/B for whether the rotation is earning its keep. No effect when follow_taps is 0. |
| isolate_references | BOOLEAN | true | Stop the text and reference rows attending to the video being denoised. Nothing reads their output and their timestep never moves, so reading the noisy target is the only reason they drift at all — cutting it makes them a pure function of the prompt and the references. On a regenerate pass the prefix is around a third of the sequence and this is worth more than any window setting: 77% of blocks kept becomes 54%. Does what the separate Reference Isolate node does, which cannot run alongside this one. Target audio is untouched and keeps reading the video, or the soundtrack stops matching the picture. |
| start_percent | FLOAT | 0.000–1 | Denoise fraction the window switches on at. Above 0 the early steps run dense — those are the ones that decide global composition and motion, so raise this first if the clip loses its overall shape rather than its detail. |
| end_percent | FLOAT | 1.000–1 | Denoise fraction it switches off at. 1.0 keeps it on to the end. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |