LTX25 Spatial Split Params

This is the inner loop of LTX25 Ultimate Upscale: it decides how each time chunk gets cut into spatial tiles so the sampler never has to hold a whole frame in VRAM at once. Its spatial_split_param output wires into the main node's input of the same name. Don't connect it and each chunk is sampled whole - which is fine for small frames, and pointless for the high-res upscaled output this pack exists to produce.

The two tile-sizing modes

specific_size (default) is the direct one: set tile_width/tile_height in pixels. Both must be multiples of 32 - one LTX VAE token. Defaults are 512x512, which is a sane starting point on a mid-range card.

rows_cols is the smarter one for a known target: give it upscale_width/upscale_height (the upscaled frame size, matching what you set in LTX25 Latent Upscale Params) plus grid_rows x grid_cols (1-9 each), and it solves an equal-size tile grid for you - every tile identical, edge tiles included - with the per-axis overlap auto-solved to a multiple of 32. If the solved tiles would be smaller than min_tile_size, it errors rather than silently giving you slivers. Bonus: the node exposes the resolved tile_width and tile_height as real INT outputs, which is handy if you want to wire the actual tile size into other nodes.

The seam controls

• spatial_w_overlap / spatial_h_overlap (default 128px) - overlap between neighbouring tiles, multiples of 32 and smaller than the tile. In rows_cols mode these are desired values; the node solves the real ones.
• fade_width / fade_height (default 32px) - the width of the fade segment at the interior edge of the overlap band. The band splits into a frozen segment (seam side, mask 0) plus this fade. Set to 0 to freeze the entire overlap strip; it clamps to the solved overlap in rows_cols mode.
• min_tile_size (default 256) - edge tiles smaller than this get pulled back until they qualify.
• overlap_mode - who wins a shared band on stitch: earlier (already-stitched content wins, default) or later (the re-sampled tile wins).
• overlap_blend - how the band transitions: linear cross-fade (default), smoothstep, overwrite, or midpoint.

The economics, briefly

Tiling isn't free. Each tile is sampled at its true extent plus overlap, so the redundancy is roughly ((tile + overlap) / tile)² - 512px tiles with 128px overlap is about 1.56x wasted compute. Shrink tiles to 256px with the same 128px overlap and it jumps to ~2.25x. So the sweet spot is the largest tiles that keep peak VRAM just under capacity; the README's guidance to keep tiles at 256px or above applies to LTX2.5's 32px grid too. Watch the VRAM meter during the first tile and adjust one axis at a time.

A note on what the frozen-overlap machinery buys you: each tile's shared seams are pre-filled from already-stitched neighbours and locked with a noise mask, so the re-sample can only change the free interior. That's what stops tiled upscaling from looking like a patchwork quilt - the seam content is preserved, not re-guessed.

Installing

Part of bbaudio-2025/Comfyui-MMH3-UltimateUpscale. ComfyUI Manager (search "Comfyui-MMH3-UltimateUpscale"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/bbaudio-2025/Comfyui-MMH3-UltimateUpscale

Restart. No pip deps; needs a current ComfyUI.

Troubleshooting

• "must be a multiple of 32 pixels" - anything off the LTX grid errors at execution. Keep tiles, overlaps, fades and min_tile_size on multiples of 32.
• "spatial_w_overlap must be smaller than tile_width" / "fade_width must not exceed spatial_w_overlap" - the validation is explicit; read the error and fix the offending pair.
• "solved tile size … smaller than min_tile_size" (rows_cols) - reduce grid_rows/grid_cols, or lower min_tile_size.
• Blocky tile boundaries - raise the overlap, or switch overlap_blend to smoothstep; if tiles are diverging, the freeze-mask logic is doing its job and your problem is per-tile sampling consistency, not the split.