Skip to content

Mask Cleanup

The first stop for a mask from SAM, rembg or any other cutout tool. It deletes stray specks, fills small holes inside the subject, and can finish with an even feather.

Mask Cleanup — before / after
In ComfyUI Add Node → ComfyUI-JH-PixelPro → mask → Mask: Cleanup
Good for Removing floating dots, filling pin-holes in a subject, producing a clean evenly feathered mask

When to use it

AI segmentation often leaves two kinds of mess: tiny floating islands of selection in the background, and small holes inside the subject. Both come from the mask being a set of connected pixel groups. This node finds each group and decides by its size in pixels whether to keep it, instead of by a brush radius. That makes it more predictable than erode-and-dilate for tidying up.

The mask is first turned into pure black and white at the threshold. Then islands smaller than min_island_area are removed, and holes smaller than max_hole_area are filled. A hole that touches the image border is never filled, because it is really background.

Quick start

  1. Connect the mask from your segmentation node.
  2. Keep the defaults (island 64, hole 64, no feather). Raise the areas for bigger images.
  3. Add feather_px 3 to 6 if you want a soft edge.
  4. Compare before and after with Mask Preview.

Settings

Setting Default What it does Tip
threshold 0.5 Pixels above this value count as subject. Lower it to include faint areas.
min_island_area 64 Delete subject blobs smaller than this many pixels. 0 = off. Scale with image size. Try 200 to 2000 on a 2 K photo.
max_hole_area 64 Fill enclosed holes smaller than this many pixels. 0 = off. Never fills background that touches the image border.
feather_px 0.0 Soft-edge distance, measured from the cleaned edge. 0 = hard black and white. 0 to 256. The full soft ramp spans twice this value.
connectivity 8 8 treats diagonal pixels as connected; 4 only counts side neighbours. Leave on 8 for most masks.

Output: mask (MASK).

Tips from retouching

  • Use it first in a mask chain, then Edge-Aware Mask Refiner or Sub-Pixel Mask Refiner.
  • Its feather is an even distance from the edge, so the half-way line stays on the original outline. A blur would shrink thin parts of the subject instead.
  • If a real part of the subject disappears (a small earring, a button), min_island_area is too high.
  • If a real hole disappears (a gap between an arm and the body), lower max_hole_area.
Under the hood

The mask is binarised with mask > threshold. Components are labelled with cv2.connectedComponentsWithStats. Holes are found on the inverted mask using the complementary connectivity (8 for the subject pairs with 4 for the holes), which is the standard pairing that stops a hole from leaking diagonally. The feather is alpha = clip(0.5 + sd / (2 * feather_px), 0, 1), where sd is the signed distance to the boundary from cv2.distanceTransform (positive inside). It runs on the CPU, one batch item at a time. Source: core/mask_cleanup.py.

Example workflow

Mask Cleanup example workflow. Copy workflows/sample_portrait.jpg and workflows/sample_binary_mask.png into ComfyUI/input/ first.