Skip to content

Sub-Pixel Mask Refiner

Takes a jagged, hard-edged cutout mask and gives it a smooth, feathered edge, while keeping the middle of the subject fully opaque and the background fully clear.

Sub-Pixel Mask Refiner — before / after
In ComfyUI Add Node → ComfyUI-JH-PixelPro → mask → Sub-Pixel Mask Refiner
Good for Softening SAM or rembg cutouts, removing the "sticker" look, preparing a mask for compositing

When to use it

A binary mask has only two values: 0 (not selected) and 1 (selected). Every edge pixel is either in or out, so the edge looks stair-stepped and pasted on. A soft mask (also called an alpha) has in-between values along the edge, which blends the subject into the new background.

A plain blur would soften the edge, but it would also eat into the subject and let the background bleed in. This node is smarter: it marks a safe zone inside the subject and a safe zone outside it, and only feathers the thin band between them. Use it when your mask is already roughly in the right place and you only need a better edge. If the edge has to follow fine detail like hair, use Edge-Aware Mask Refiner or Alpha Matte instead.

Quick start

  1. Load Image → Convert Image to Mask (or any node that outputs a mask).
  2. Connect the mask to Sub-Pixel Mask Refiner.
  3. Start with the defaults (erosion 2, dilation 4, feather 2.0). Raise feather_sigma for a softer edge.
  4. Check the result with Mask Preview.

Settings

Setting Default What it does Tip
erosion_radius 2 How far to pull back from the edge to find the "definitely inside" core. That core stays fully white. 0 to 64. Bigger = a wider feathered band on the inside.
dilation_radius 4 How far to push out from the edge to find the "definitely outside" zone. That zone stays fully black. Keep it equal to or larger than erosion_radius.
feather_sigma 2.0 Softness of the edge, in pixels. 0.1 to 32. Try 1 to 4 for portraits.
threshold 0.5 Values above this count as selected before anything else happens. Only matters if your input is not already pure black and white.

Output: refined_mask (MASK), same size as the input.

Tips from retouching

  • Feed it a clean binary mask. If the mask has specks or holes, run Mask Cleanup first.
  • Set dilation_radius a little larger than erosion_radius so the soft band is centred on the original edge.
  • Setting both radii to 0 turns the node into a plain Gaussian blur of the mask. That is rarely what you want.
  • For hair and fur, a feather alone is not enough. Build a Trimap and use Alpha Matte.
Under the hood

The mask is binarised with mask > threshold. A Gaussian blur of the binary mask gives the feather. An eroded copy is the inside core and 1 - dilated copy is the outside core. The output is inside + (1 - inside - outside) * feather, clamped to 0..1, so the feather only shows in the band between the two cores.

Erosion and dilation use a square window (a sliding max, done with separable max_pool2d), not a round one, so memory stays proportional to the image size even at large radii. The Gaussian kernel is sized to cover about three sigma. A warning is logged when more than 5% of the input pixels are mid-grey, because the node expects a near-binary mask. Source: core/mask_refiner.py.

Example workflow

Sub-Pixel Mask Refiner example workflow. Copy workflows/sample_binary_mask.png into ComfyUI/input/ first.