Mpi Windowed Sampler

Video models have a nasty habit: your settings work at one clip length, you add ten frames, and suddenly the card dies. Not bad luck - a video DiT's attention cost scales with T×H×W, so the same sampler that succeeds at one duration OOMs at another with nothing else changed. MpiWindowedSampler is a SamplerCustomAdvanced that samples over overlapping temporal windows and cross-fades the results back together, so a clip too long to refine in one pass still fits. It ships in ComfyUi-MpiNodes, Mad Pony Interactive's big utility pack (the folks behind the Cubric Vision app).

The numbers that motivated it: on MiniMax H3 at 2K on a 16GB card, T=27 samples fine while T=32 and T=37 both die inside comfy_kitchen.prequantize_int8_attention. H3's trained minimum is 124 frames, which is T=37 - the model's shortest legitimate clip is exactly the one a 16GB card can't refine. Lowering the upscale factor doesn't help: stage-2 peak memory is set by the output latent size, not the factor.

The one rule: this is a refine pass

Windowing is unsound for first-pass generation - cut a cold generation into windows and each one will decide its own motion, identity, and framing, and the seams will fight. It works for a refine precisely because the latent arrives already coherent: motion and framing were decided upstream, a short sigma schedule only adds local detail, and two windows have nothing to disagree about. So use it where a hi-res-fix stage 2 goes - a stage-1 latent upscaled to target, low sigmas, adding detail. First-pass through it and you'll get ghosts at every seam.

How the seams disappear

Three decisions make the cross-fade work. Noise is generated once for the whole clip and sliced per window, so both sides of a seam denoise identical noise over identical input - the fade blends two near-identical results instead of two different ones. Windows are cut on the model's temporal grid - H3 packs time as a 2-frame causal head plus blocks of 5, so a legal latent length is 5k+2 and a legal cut point is a multiple of 5; a window ending four frames into a block gets padded and decodes black, an artefact that reads as a seam problem and isn't one. The model stays loaded across windows on purpose: unloading between them would re-stream the whole DiT over PCIe per window, costing more than the sampling it protects.

On an audio-video model (H3, LTXV), the audio in the joint AV latent is sliced to its window for conditioning and restored whole on output - hand a window the entire soundtrack and it thinks frame 15 begins at t=0, so a mouth starts moving two seconds late. The refined audio is discarded and the original stream kept, so nothing stitches a seam into the sound.

The inputs that matter

Wire it exactly like SamplerCustomAdvanced - noise (NOISE), guider (GUIDER), sampler (SAMPLER), sigmas (SIGMAS), latent_image (LATENT) - then three widgets. Both frame widgets count in video frames - the unit an OOM is discovered in; the model's own temporal downscale ratio plus frame_grid do the conversion.

• window_frames (default 90) - a ceiling on the video frames refined at once, not a target. If 124 dies and 90 survives, put 90 here. Windows then shrink to the smallest legal size that covers the clip in the same pass count, so a generous ceiling costs nothing.
• overlap_frames (default 17) - the minimum video frames neighbouring windows share and cross-fade. It costs compute, but skimping invites ghosting: on H3 at T=37, one measured seam went from a face visibly splitting in two, to a trace of distortion at 7 shared frames, to clean at 17 - and 17 is one whole H3 block.
• frame_grid (default 5) - advanced. The model's temporal quantum in latent frames. Leave 5 for H3; set 1 for a model with no temporal packing, which gives plain fixed-size windows. Snapping can only ever hand you more overlap than you asked, never less.

Outputs

• output (LATENT) - the refined, cross-faded latent, audio restored whole.
• denoised_output (LATENT) - the blended x0 prediction, accumulated separately (some graphs, like LTX i2v/t2v, wire it).
• info (STRING) - a report of what was planned: window count, real shared frames, grid, in video and latent units (also printed to the console). Read it once and the grid snapping stops being a mystery.

A clip already at or under the ceiling skips all of this and takes a plain single-pass path, unchanged.

Install & troubleshooting

Same as every node in this pack: install via ComfyUI Manager (search "ComfyUi-MpiNodes"), or manually and restart:

cd ComfyUI/custom_nodes && git clone https://github.com/MadPonyInteractive/ComfyUi-MpiNodes

No models, no compiled deps - it's all ComfyUI's own sampling API.

• Black frames or flashing at the tail → off-grid window. Keep frame_grid at 5 for H3; an over-ceiling window risks an OOM, an off-grid one guarantees corruption.
• Seam ghosting, a face that doubles → not enough overlap. Raise overlap_frames - the cross-fade needs to be wide enough to hide two valid-but-different denoisings.
• Still OOM → set window_frames from your own card's OOM discovery (124 died, 90 survived → 90). That's the feature.