🎭 Gemini Mask Stitch

Gemini Mask Stitch is the "paste back" half of the automated detailing loop - the counterpart to Gemini Mask Crop. Mask Crop cut a region out; this node takes the refined version of that region and composites it onto the original at exactly the right place, with a seam that's actually invisible. It's local, no API, no key, and it's where the quality of a whole refinement pipeline gets decided. A perfect inpaint that's pasted back with a hard edge reads as a bad edit; this node exists to stop that.

The KB's detailing loop is detect → crop → upscale → re-render → paste, and every one of those steps has a Gemini node in this pack except the paste - until now. Mask Stitch closes it.

How it works

Three inputs: original_image (the full-resolution base, which becomes the bottom layer), refined_image (your inpainted/edited region), and mask (where to place it - if you feed a batch of masks they collapse via amax into a single union alpha). The refined image is sized into the mask's bounding box according to fit_mode:

• cover (default) - preserve aspect, scale to fill the bbox, crop the overflow. Best for Nano Banana inpaint output, which tends to fill its frame.
• contain - preserve aspect, fit inside the bbox; gaps reveal the bottom layer underneath.
• stretch - fill the bbox exactly, distorting aspect. The legacy behavior; avoid unless you know the aspect already matches.

Then the seam work happens. pre_shrink_alpha (default 2) erodes the alpha by a couple of pixels before anything else, which kills the classic 1-pixel halo of original color that bleeds at the boundary - the author's description is "kills the 1-px halo," and it does. feather_pixels (default 8) then softens the alpha edge so the transition is gradual. And match_color runs a Reinhard mean/std color transfer from the underlying original region onto the refined patch, which fixes the "this patch looks brighter" tell that plagues cloud-refined regions with slightly different white balance.

Outputs

• composited - the finished merge.
• top_layer and bottom_layer - the two layers separately, so you can re-blend or debug non-destructively.
• alpha - the actual alpha used, so you can inspect exactly where the blend happened.

Having the layers and alpha out is the genuinely nice part: if the seam still shows, you don't have to rerun the pipeline - grab the alpha, shrink it with Gemini Mask Resize, and recombine.

Wiring it up

The intended chain is right in the node names: Gemini Segmentation or Object Detection → Gemini Mask Crop (images_list/alphas_list) → Gemini Batch Refine → Gemini Mask Stitch. Wire Mask Crop's alphas_list into the mask input, Batch Refine's refined_images into refined_image, the original into original_image, and the loop closes.

Install

Standard for the pack: ComfyUI Manager search "ComfyUI_Gemini", or git clone https://github.com/aarnoatchi/ComfyUI_Gemini into custom_nodes and pip install -r requirements.txt. No scipy, no API key - pure torch on the input device. If your seam shows a colored edge, that's pre_shrink_alpha too low; if it looks like a soft smear, feather_pixels too high. And if the patch is squashed, your refined_image came from a stretch-oriented crop path - switch fit_mode to cover and it'll stop distorting.