Mpi Step Windowed Sampler

Generates a clip in overlapping temporal windows, blending the windows after every sampling step so they settle on one video. Works from pure noise (first-pass generation), unlike Mpi Windowed Sampler, which is a refine. Video and audio are both windowed. Needs a stateless sampler (euler). A clip that fits one window samples in a single ordinary pass. Optional window_guider_N gives window N its own guider (its own slice of the prompt); an unwired one uses the main guider.

By MadPonyInteractive·Created about a year ago·Updated a day ago· 3

noise
guider
sampler
sigmas
latent_image
window_guider_1
window_guider_2
window_guider_3
window_guider_4
window_guider_5
window_guider_6
window_guider_7
window_guider_8

output
denoised_output
info

◄window_frames0►

◄overlap_frames17►

◄frame_grid5►

◄commit_steps0►

CategoryMpiNodes/Sampling

Inputs (17)

Name	Type	Default	Description
noise	NOISE		—
guider	GUIDER		—
sampler	SAMPLER		—
sigmas	SIGMAS		—
latent_image	LATENT		—
window_frames	INT	00–100000	CEILING on the VIDEO frames refined at once, not a target. 0 = AUTO: sized from the engine GPU's usable VRAM and picture size: the fastest window the card holds, from MiniMax H3 measurements. Otherwise this is the number you discover by hitting an OOM: if 124 frames dies and 90 survives, put 90 here. Windows are then shrunk to the smallest legal size that still covers the clip in the same number of passes, so a generous ceiling costs nothing and a low one costs passes. A clip at or under this length takes the plain single-pass path.
overlap_frames	INT	170–100000	MINIMUM VIDEO frames neighbouring windows share and cross-fade. Costs compute on every seam. 17 is one whole block for H3. Snapping to the grid can only ever hand you more than you ask for, never less.
frame_grid	INT	51–64	ADVANCED - the model's temporal quantum, in LATENT frames. Leave it at 5 for MiniMax H3, which packs time as a 2-frame causal head plus blocks of 5; a window off that grid ends in a padded part-block that decodes BLACK. Set 1 for a model with no such structure, which gives plain fixed-size windows and makes the two widgets above count latent frames instead.
window_guider_1opt	GUIDER		Guider for window 1 only (counted from the start of the clip), so that window hears its own slice of the prompt. Unwired = the main guider. Must wrap the same model. Ignored on the single-pass path.
window_guider_2opt	GUIDER		Guider for window 2 only (counted from the start of the clip), so that window hears its own slice of the prompt. Unwired = the main guider. Must wrap the same model. Ignored on the single-pass path.
window_guider_3opt	GUIDER		Guider for window 3 only (counted from the start of the clip), so that window hears its own slice of the prompt. Unwired = the main guider. Must wrap the same model. Ignored on the single-pass path.
window_guider_4opt	GUIDER		Guider for window 4 only (counted from the start of the clip), so that window hears its own slice of the prompt. Unwired = the main guider. Must wrap the same model. Ignored on the single-pass path.
window_guider_5opt	GUIDER		Guider for window 5 only (counted from the start of the clip), so that window hears its own slice of the prompt. Unwired = the main guider. Must wrap the same model. Ignored on the single-pass path.
window_guider_6opt	GUIDER		Guider for window 6 only (counted from the start of the clip), so that window hears its own slice of the prompt. Unwired = the main guider. Must wrap the same model. Ignored on the single-pass path.
window_guider_7opt	GUIDER		Guider for window 7 only (counted from the start of the clip), so that window hears its own slice of the prompt. Unwired = the main guider. Must wrap the same model. Ignored on the single-pass path.
window_guider_8opt	GUIDER		Guider for window 8 only (counted from the start of the clip), so that window hears its own slice of the prompt. Unwired = the main guider. Must wrap the same model. Ignored on the single-pass path.
commit_stepsopt	INT	00–1000	For the first N steps each overlap takes the EARLIER window's prediction only (no cross-fade), so the layout is set by one window; the later window joins on the remaining detail steps. 0 = cross-fade on every step.

Outputs (3)

Name	Type	Description
output	LATENT	—
denoised_output	LATENT	—
info	STRING	—