Framed Mask Grab Cut 2

Framed Mask Grab Cut 2 is the two-input version of the pack's GrabCut node. Where the original feeds one threshold image and two threshold values, this one takes two separate images - one that says "probably foreground," one that says "definitely foreground" - and lets OpenCV's GrabCut refine from there. It's the same framed-background trick, with more explicit control over the seeds.

Why you'd reach for it

GrabCut's quality lives or dies on its seed flags. Version 1 gets there by thresholding a single image two ways; version 2 lets you supply the two seed sources yourself, which is exactly the situation you're in when one mask is a confident detection and another is a fuzzy guess. Feed the fuzzy one to thresh_maybe and the confident one to thresh_sure, and GrabCut gets the layered hint it thrives on.

The real niche is when you have a two-tier signal available - say, a strict segmentation output (sure) and a loose bounding-region mask (maybe). That's a common pairing in real CV pipelines, and this node is built for it.

How it works

Both thresh_maybe and thresh_sure are thresholded against the same binary_threshold (default 128) - it's the safeguard for inputs that aren't clean 0/255 images. Pixels >= binary_threshold in thresh_maybe become probable foreground (GC_PR_FGD); pixels >= binary_threshold in thresh_sure become sure foreground (GC_FGD) on top of that. The frame - margin pixels on the border - is painted sure background (GC_BGD), with frame_option to skip sides (IGNORE_TOP/BOTTOM/LEFT/RIGHT/HORIZONTAL/VERTICAL).

Then there's the option the first version doesn't have: maybe_black_is_sure_background. Flip it on and pixels that are black in thresh_maybe become sure background rather than just probable - a strong claim that anything your loose mask leaves out is genuinely background. Default is off, because that claim isn't always true.

GrabCut runs iterations passes, and the output is a mask (background/probable-background → black, everything else → white) in RGB or GRAY per output_format.

The inputs that matter

• thresh_maybe / thresh_sure - your two seed images, both thresholded at binary_threshold.
• maybe_black_is_sure_background - the big new lever. On when your maybe-mask is a reliable enclosure; off when it's noisy.
• iterations - refinement passes (default 25).
• margin / frame_option - size and sides of the guaranteed-background frame.

Output is a single IMAGE mask.

Install

OpenCV-backed, so requirements matter:

cd ComfyUI/custom_nodes
git clone https://github.com/bmad4ever/comfyui_bmad_nodes
pip install -r requirements.txt

then restart ComfyUI. Manager users search "comfyui_bmad_nodes"; without OpenCV the CV nodes won't load.

Common issues

• Both inputs look the same? Then you're not actually giving GrabCut a two-tier hint - use v1, which is simpler for single-source cases.
• Background bleeding in. Turn on maybe_black_is_sure_background if your maybe-mask reliably excludes the background.
• Subject touching the frame. Exclude that side with frame_option, since the frame is declared background.
• Zero iterations? iterations can be 0, which means "seed mask only, no refinement" - occasionally useful for debugging, rarely what you want.

Version 2 is the more capable sibling, but only if you actually have two distinct seed sources. If you have one threshold image, the original Framed Mask Grab Cut is the right tool.