Image Untile

If you've ever tried to upscale a 4K image on a 6GB card, you know the drill: you tile it, process the tiles, and then stare at the seams. wcx_ImageUntile is the "put it back together" node for this pack's Image Tile, and the reason you'd reach for it is exactly the tiled-diffusion situation the KB's upscaling doc describes - big output, small VRAM, tiles processed one at a time.

The mechanism is straightforward but you have to feed it the right numbers. It assumes the tile batch arrived in row-major order - tile 0 is the top-left, then left-to-right across the columns, then down the rows - which is exactly how Image Tile emits them. It works out the base tile size by subtracting the overlap you tell it from the first tile's dimensions, builds an empty canvas of rows × columns tiles, and places each one. Where tiles overlap, it feathers the shared edge with a linear alpha ramp instead of just hard-stacking them, which is what hides most of the grid.

Inputs are the four numbers that describe your tiling setup:

• Tile Batch (切片批次) - the stack of tiles. Note the port label shows up in Chinese in the UI; it's the IMAGE batch.
• Overlap Horizontal (重叠横向) and Overlap Vertical (重叠纵向) - how many pixels the tiles shared in each direction.
• Rows (行数) and Columns (列数) - the grid you tiled into.

One IMAGE comes out, ready for a preview or save node.

The traps are the usual suspects for stitching:

• You must feed back the same rows, columns, and overlap you tiled with. Get any of them wrong and the image comes out scrambled or with doubled regions. There's no sanity check - it just trusts you.
• The feather is a linear ramp, not a seam solver. It blends, but it can't fix tiles that genuinely diverged. If your tiles drifted because the upscale pass ignored its condition, the grid lines show through no matter how much feather you add. That's the KB's warning about skipping ControlNet Tile on tiled upscales - the condition is what keeps tiles faithful, the untile just hides the edges.
• Tiles must be uniform size. The node computes dimensions from the first tile and assumes the rest match. If something upstream produced ragged tiles, expect misalignment.

Install is the standard two paths, and the pack has zero heavy dependencies - no model files, no requirements file, just Python.

# ComfyUI Manager: search "ComfyUI-Practical-Tools" → Install → restart

# or manually:
cd ComfyUI/custom_nodes
git clone https://github.com/wenchengxiang/ComfyUI-Practical-Tools
# restart ComfyUI

It's a small, honest node. The only thing to read before you trust it is this: same grid, same overlap, row-major order.