Image Crop Merge

Crop-and-paste workflows have a silent killer: after you crop a region, run it through some node, and try to paste it back, the paste rarely lines up. UC_ImageCropMerge exists so you don't hand-compute the paste math - give it the processed crop, the original image, and the crop's coordinates, and it resizes the crop to fit and merges it back into the right spot. This is the "put it back" half of a face-detail or region-inpaint loop.

It pairs naturally with the pack's bounding-box and crop nodes: crop a face region, upscale or inpaint it, then merge the result back over the original at the same coordinates. The resize-to-region behavior is the key feature - your processed crop can come back at a different resolution than the original crop was, and the node fits it to the destination region so you don't have to.

How it works

cropped_image is resized to the destination region defined by crop_x, crop_y, crop_width, crop_height (all in original-image pixels), using resize_method (default lanczos), and pasted into original_image at that spot. An optional mask soft-blends the merge; leave it disconnected and the region is replaced completely.

The inputs that matter

• cropped_image - the processed crop.
• original_image - the full image receiving it.
• crop_x / crop_y / crop_width / crop_height - the region in original-image pixels. These are force_input sockets, so they're meant to be wired from whatever computed the crop - wire them from the same crop/bounding-box node that produced the crop and they'll always agree.
• mask - optional blend mask for the region.

Output: a single IMAGE with the merged result.

Installing it

ComfyUI Manager → search "ComfyUI-UtilsCollection", or:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-UtilsCollection

Restart. Deps: opencv-python and typing-extensions, no models. Newer-API pack - keep ComfyUI updated.

Where it bites: stale coordinates. If the crop node and this node get their crop_x/crop_y from different sources - or one of them is a hardcoded widget - the region drifts and the "fixed" face lands half a chin off. Wire the coordinates from the same socket everywhere. Also remember: it resizes the crop to the region, so an extreme aspect mismatch between them will distort the crop; keep the destination region near the crop's aspect ratio unless distortion is the point.