Split Image Grid

split chops one image into a grid of equal tiles and hands them back as a batch. You tell it how many columns and rows, and it slices accordingly - a 2x2 gives you four quarters, a 3x1 gives you three vertical strips. Its natural partner is Merge Image Grid, which puts them back together.

Why you'd reach for it

Two main jobs. The obvious one is un-gridding: a lot of models and workflows produce contact-sheet style outputs - a batch laid out as a single grid image - and you want the individual frames back. split reverses that in one node.

The other, more interesting job is manual tiling. Diffusion at high resolution is VRAM-limited, and the standard escape is to break the image into tiles, process each one, and stitch them back - that's the whole idea behind tiled upscaling. split lets you build that by hand: cut the image into a grid, run each tile through whatever you like (an upscale pass, a filter, a per-tile effect), then reassemble with merge. Be honest with yourself about what this is, though - it's a plain grid split with hard tile boundaries, not the overlapping-tile-with-blending machinery that Ultimate SD Upscale or Tiled Diffusion give you. For serious seamless upscaling those purpose-built nodes are still the right call. split/merge shine when you want simple, predictable tiling you fully control.

How it works

It divides the image into columns x rows equal cells and outputs them as a list of images, read in grid order. Because the output is a list, every downstream node runs once per tile automatically - that's ComfyUI's list behavior, and it's what makes "process each tile" just work without a manual loop.

The inputs and outputs

• image - the image to cut up.
• columns - how many columns in the grid (default 1).
• rows - how many rows (default 1).

The output is image, and importantly it's a list - one entry per tile. Feed it into your per-tile processing, then collect the results back into Merge Image Grid with the same columns and rows to rebuild the full picture.

Installing it

ComfyUI Manager → search antrobots ComfyUI Nodepack, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/antrobot1234/antrobots-comfyUI-nodepack

No models or dependencies to download.

Common issues

The thing to keep straight is that the output is a list, so everything after it fans out to run per-tile. That's usually what you want, but it means a 4x4 split turns one sampling pass into sixteen - mind the time and cost if the per-tile work is heavy.

The classic seam problem: because tiles are cut on hard boundaries with no overlap, any generative work you do per-tile can produce visible edges where they meet after merge. That's inherent to naive grid tiling, not a bug in the node. If you're seeing seams and you need them gone, that's the signal to reach for an overlap-and-blend upscaler instead.

Finally, keep columns and rows identical between split and merge. Mismatched grid dimensions are the usual reason a reassembled image comes out scrambled or the wrong shape.