Image Untile (Seam Mask) (Yogurt Nodes)

Tiling is easy. The part that makes or breaks a crop-and-stitch workflow is the untile - because if you just paste tiles side by side, you get a quilt, with hard lines wherever two independently-generated tiles meet. Image Untile (Seam Mask) is the part that doesn't. It takes the tiles, masks, and tile_info from its sibling Image Tile (Seam Mask) and merges everything back into one image with feathered, overlap-aware seams.

It's an image node in the YogurtNodes pack (yogurt7771/ComfyUI-YogurtNodes), and it's the second half of the pack's answer to high-res generation: tile → regenerate each tile at native quality → untile with smooth transitions.

How it works

The mechanism is weighted accumulation. Each tile gets laid onto a blank canvas at the position recorded in tile_info, but instead of overwriting, the node accumulates pixels and a weight map. In the overlap regions, tiles blend according to a seam-feather weight that fades each tile's contribution toward its edges - so where two tiles overlap, the transition is a gradual crossfade, not a hard cut. The final image is accumulated_pixels / accumulated_weights, a proper normalized blend.

The masks aren't ignored either: they influence tile priority, so a tile that was actually regenerated (white = inpaint) contributes more strongly where it should, while reference regions (black) keep the original pixels stable. The comment in the source is worth quoting in spirit: the seam feather is always in effect, and the mask only increases a tile's priority. That means even if you feed masks that are off, you still get smooth seams.

Inputs that matter

• tiles - the tile batch from Image Tile (Seam Mask).
• masks - the matching mask batch (white = inpaint, black = reference).
• tile_info - the dict from the tile node. This is the one you must not lose; it carries tile positions, the original canvas size, and the feather settings. Losing it is the #1 way to break this node.

Output: a single image, back at the original size.

Install

One clone, whole pack:

cd ComfyUI/custom_nodes
git clone https://github.com/yogurt7771/ComfyUI-YogurtNodes.git
cd ComfyUI-YogurtNodes
pip install -r requirements.txt

Or ComfyUI Manager → "YogurtNodes". Find it under Yogurt Nodes / Image. No models, no downloads.

Troubleshooting

• "Tiles must be square" / "tile_size mismatch." The node requires square tiles whose edge matches tile_info.tile_size. If you resized the tiles between tile and untile (say, to speed up the inpaint), that's the error - resize back to tile_size before merging, or re-tile.
• "Batch size not divisible by tile_count." The tiles batch has to be an exact multiple of the tile count from tile_info. A dropped tile somewhere in your graph breaks the arithmetic. Check that no tile got filtered out between the two nodes.
• masks batch mismatch. If your masks don't match the tiles batch, the node tries to broadcast a single mask - fine if you have one, an error if the counts disagree.
• Single tile? If tile_info says one tile, it returns the tiles as-is. Good for testing.

The workflow to remember: Tile → inpaint/regenerate each tile (respecting the masks - white gets redrawn, black is context) → Untile with the same tile_info. Keep the tile/mask/untile trio bundled and the seams just disappear. It's the closest thing to a plug-and-play high-res pipeline in the pack.