MMH3 Spatial Split Params

This is the node that decides whether your upscale runs on your GPU or dies in a fireball of "out of memory". MMH3 Spatial SplitParams bundles the tile settings for MMH3 Ultimate Upscale: how each video frame gets chopped into overlapping tiles, how the seams between tiles get stitched, and - the part that matters most to you - how big a single tile is. Because peak VRAM is bounded to one tile, tile size is your VRAM budget. This is tiled diffusion for H3, and this node is the dial.

The mechanism in plain terms

Each chunk (from the temporal split) is cut into a grid of overlapping tiles. One tile is sampled at a time. Here's the trick that keeps seams invisible: the overlap strip a tile shares with an already-stitched neighbor is split into two segments. The part nearest the seam is frozen - pre-filled from the accumulated result and locked with a noise mask, so the re-sample literally cannot change it. The part nearest the tile's interior is the fade segment, a smooth mask transition. After sampling, the frozen seam is written back with torch.where, guaranteeing the already-consistent content is never overwritten. That's how you get big upscales without the blocky checkerboard that naive tiling produces.

The inputs that matter

Two of them do 90% of the work, and the rest are seam polish:

• tile_width / tile_height (default 512) - tile size in pixels at the upscaled chunk resolution, multiples of 32. Smaller = less VRAM, more tiles, slower. 512 is a good starting point; drop to 256 or 384 on a small card.
• spatial_w_overlap / spatial_h_overlap (default 128) - overlap in pixels between neighbors, also multiples of 32 and must be smaller than the tile size. More overlap = better consistency, more wasted work. 128 is the sensible default.

Then the seam controls: fade_width / fade_height (default 32) set the length of the fade segment; set to 0 to freeze the entire overlap strip. min_tile_size (default 256) pulls the last edge tile back so it never degenerates into a sliver. And overlap_mode (earlier or later - who wins the shared band) plus overlap_blend (linear / smoothstep / overwrite / midpoint) control exactly how adjacent tiles transition. The defaults - linear, earlier - are the safe ones; the others exist for when a seam shows.

The single output, spatial_split_param, plugs into the spatial_split_param input on MMH3 Ultimate Upscale. Don't connect it and the node samples each chunk whole, which is fine for short clips at native-ish size.

Install and troubleshooting

Same pack, same install: Comfyui-MMH3-UltimateUpscale in ComfyUI Manager, or clone the repo into custom_nodes and restart. No dependencies, no models - this node is pure geometry.

Real failure modes to expect, all with the author's own validation baked in:

• "must be multiples of 32" - the node checks this and throws a ValueError. Tile dims, overlaps, and fades all have to land on the 32px grid.
• "overlap must be smaller than tile" - obvious in hindsight, but easy to hit while dialing in a small tile.
• "Upscale width must be >= tile_width" - if you're tiling and upscaling, the upscale target has to be at least as big as the tile, or the grid makes no sense.

If tiles are too small you'll get more seams to manage and slower runs for marginal VRAM savings; too big and you've recreated the OOM you were trying to dodge. The overlap/fade defaults are well chosen - tweak tile size first, leave the seam math alone until you can see a problem.