Experimental per-step spatial Tiled Diffusion for MiniMax H3 (MultiDiffusion / Mixture of Diffusers style, ported from shiimizu/ComfyUI-TiledDiffusion). Patches the model so every denoising step slices the latent into spatial tiles, runs the DiT per tile, and blends the predictions with uniform or gaussian weights - all tiles share the same sigmas and noise field, so no overlap masks are needed. Tiles carry global frame positions so their predictions stay consistent with the full frame. Connect between the H3 model and 'MMH3 Ultimate Upscale' and leave that node's spatial_split_param unconnected. Use a tile_overlap of at least ~25% of the tile size, or the tiles decouple. Does not support fun controlnet, inpaint or denoise masks, and does not save VRAM/time (H3 runs batch 1); it is an experiment in seam quality.
By bbaudio-2025·Created about a month ago·Updated a day ago· 208
MMH3 Tiled Diffusion (Experimental)
model
model
◄methodgaussian►
◄tile_width1024►
◄tile_height1024►
◄tile_overlap256►
Categorymodel/latent/minimax
Inputs (5)
Name
Type
Default
Description
model
MODEL
The MiniMax H3 diffusion model to patch.
method
COMBO
gaussian
Overlap blending weight per tile: gaussian (Mixture of Diffusers, default - tiles dominate their center) or uniform (MultiDiffusion - plain averaging in overlaps).
tile_width
INT
102464–8192
Tile width in PIXELS, snapped to a multiple of 32 (the H3 2x2 latent patch grid).
tile_height
INT
102464–8192
Tile height in PIXELS, snapped to a multiple of 32 (the H3 2x2 latent patch grid).
tile_overlap
INT
2560–8192
Overlap between neighbouring tiles in PIXELS, snapped to a multiple of 32. Keep at least ~25% of the tile size or the tiles decouple into unrelated content. 0 gives hard tile edges.
Outputs (1)
Name
Type
Description
model
MODEL
The patched model; feed it into 'MMH3 Ultimate Upscale' (spatial_split_param unconnected) or any H3 sampler.
