GrabCut Mask Refinement

Sometimes you already have a mask - from RMBG, BiRefNet, a manual paint job - and it's close but wrong at the edges: bleeding into the background, eating into the subject, fuzzy around contours. GrabCut Mask Refinement exists for exactly that. You give it an IMAGE and a MASK, it uses the mask to define where the subject roughly is, then runs GrabCut to pull the boundary back onto the actual object edges using the color and contrast information in the source image. It's the classic GrabCut "refine a trimap" workflow, packaged as a node. This is the pack's honest use case for GrabCut: not segmenting from nothing, but fixing a mask that's 90% right.

Inputs that matter

• grabcut_iterations (default 3) - how hard GrabCut works; more iterations, tighter convergence
• edge_refinement (0-1) - boundary smoothing strength
• edge_blur_amount (0-10) - Gaussian blur on the mask edge; 0 keeps it sharp, higher feathers it
• expand_margin (default 10) - pixels to expand the mask boundary so GrabCut has room to work
• bbox_safety_margin and min_bbox_size - keep the refinement region sensible
• output_size (ORIGINAL or fixed presets; custom_width/custom_height for custom) and scaling_method
• invert_mask - swap foreground and background if your source mask is backwards

The README's edge_blur guidance is a decent starting point: 0.0 preserves sharp edges (pixel art), 0.5-1.5 feathers for natural photo edges, 2.0+ gives soft composite blending. With edge_blur_amount at 0 the blur path is skipped entirely, so there's no cost to leaving the knob where it is.

Two outputs: image and refined_mask. Feed refined_mask into whatever consumed the original - compositing, inpainting, a save node - and compare against the source mask to see what GrabCut actually fixed.

When it's worth it

The product-pipeline pattern, honestly: run a fast neural remover to get 90% of the mask, then snap the edges with this node. It's cheaper than re-running a heavier matting model, and it's the pack's answer when your cutout is almost there but not quite. Keep expectations calibrated - GrabCut works on color/contrast boundaries, so if the source image genuinely lacks an edge (subject and background are nearly the same tone), no amount of iteration finds one.

Install

It lives in the pack's GrabCut module, so the dependency rule from the auto node applies: ultralytics must be installed or the whole GrabCut group - this node included - won't load. Same pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/Limbicnation/ComfyUI-TransparencyBackgroundRemover.git
cd ComfyUI-TransparencyBackgroundRemover
pip install -r requirements.txt

Or search "Transparency Background Remover" in ComfyUI Manager, restart, and it'll be under image/processing.