Extensions/ComfyUI Detailed Region Inpaint
ComfyUI Extension

ComfyUI Detailed Region Inpaint

ComfyUI nodes for detailed region inpainting with smart cropping, upscaling, and seamless compositing back into the original image.

By ArtemKo7v·Created 21 days ago·Updated 20 days ago· 2
ArtemKo7v/ComfyUI-DetailedRegionInpaint
Nodes2
On cloudLocal install
CategoryArtemKo7v/inpaint
Stars2
Updated20 days ago

Nodes (2)

Prepare Inpaint Region

Give the bit you're fixing its own 1024 pixels

ArtemKo7v/inpaint
Restore Inpaint Region

The node that makes 'nothing else moved' literally true

ArtemKo7v/inpaint
Readme

ComfyUI Detailed Region Inpaint

Two ComfyUI nodes crop a masked area, enlarge it for a normal inpainting workflow, and composite the processed crop back into the source image.

Install

Place this folder in ComfyUI/custom_nodes/ and restart ComfyUI. PyTorch is supplied by ComfyUI; this package has no additional Python dependencies.

Workflow

  1. Connect the source IMAGE and MASK to Prepare Inpaint Region.
  2. Connect its enlarged image and mask outputs to your preferred inpainting nodes.
  3. Connect the resulting image to Restore Inpaint Region as inpainted_image.
  4. Also connect the original image, original mask, and Prepare's region_data output to Restore.

The two nodes are in ArtemKo7v/inpaint. The region_data socket carries the exact crop geometry, so no coordinates need to be entered manually.

Prepare inputs

| Input | Default | Meaning | | --- | ---: | --- | | padding | 32 | Context pixels around the nonzero mask bounds, clipped to the source image. | | scale | 2.0 | Requested enlargement factor; 1.0 keeps the crop at its original size unless max_size requires a reduction. | | max_size | 1024 | Maximum width or height of the prepared crop. | | mask_blur | 0.0 | Gaussian blur sigma in output mask pixels, applied after resizing. The specified value is independent of scale and max_size. |

The bounds are calculated from the original, unblurred mask. Prepare uses bicubic image resizing and bilinear mask resizing. It favors dimensions divisible by 8 when doing so keeps nearly the same aspect ratio and resolution.

Restore behavior

Restore resizes the processed image to the stored crop dimensions, then blends it into a copy of the original image using the original mask. Mask values from 0 to 1 are preserved for soft compositing; pixels where the original mask is zero remain unchanged. mask_blur affects the mask sent to the inpainting workflow, but does not feather the final composite.

The current version supports one image and one mask per call. Larger batches, empty masks, incompatible source dimensions, and invalid crop metadata raise a clear error.

Tests

Run python -m unittest discover -s tests -v from this folder. The tests cover crop bounds, edges, corners, tiny and large regions, soft masks, empty masks, metadata validation, and an unchanged-image round trip.