BD Mask Resolver

BD Mask Resolver takes a character silhouette plus overlapping category masks - skin, clothes, accessories - and produces clean, mutually-exclusive masks. Every pixel ends up in exactly one category, so downstream steps (skin repaint, part isolation, channel packing) can trust the boundaries. It's a Python port of the GLSL Mask Resolver shader this pack used in its engine pipeline, and knowing that explains both why it exists and why it behaves the way it does.

The problem it solves is real: SAM3 gives you overlapping guesses. "Skin" mask claims a hand, "clothes" mask also claims it, and nobody decided who wins. This node decides - with a priority ranking and a color-voting step, plus a neighbor-aware gap fill so the seams don't look like a jigsaw puzzle.

The inputs that matter

The three required inputs are silhouette (the universe of pixels to assign - anything outside becomes background in every output), skin_mask and clothes_mask (your overlapping candidates), and original_image (sampled per-pixel for skin-tone detection). The optional accessories_mask defaults to empty but, when wired, gets the highest default priority.

• skin_priority / clothes_priority / accessories_priority - the ranking that settles disputes. Default: accessories(3) > clothes(2) > skin(1), meaning a pixel claimed by both clothes and skin goes to clothes. Raise skin_priority above 2.0 to flip it. Every pixel gets claimed by the highest-priority mask that covers it.
• skin_color_mode - how the color vote works. adaptive_lab (default) samples the actual skin tone from your confident skin pixels and measures LAB ΔE distance per pixel, so it self-tunes across pale to dark to reddish skin. fixed_hsv is the hard-coded HSV ranges from the original shader - it fails on very pale/dark skin. both is strictest. adaptive_tolerance is the ΔE at which likelihood = 0.5 (25 is "clearly different but related"); adaptive_min_samples (50) is how many confident skin pixels are needed before adaptive mode activates.
• color_weight - how much the color vote influences scoring (0 = ignore image color, 1 = strong). The default 0.6 is a reasonable middle.
• claim_threshold - a pixel is claimed only if the winning score exceeds this. Raise it (0.5+) to make more pixels "residual" (eligible for gap fill); lower it for aggressive claiming. This is the dial that trades sharpness for coverage.
• gap_fill_strength / gap_neighbor_radius - how residual pixels get assigned via neighbor voting. Default radius 8 covers small gaps; 16–32 for high-res character art.

overlap_mode switches between priority (winner takes all - each pixel in exactly one category) and soft_blend (fractional distribution where outputs can overlap). output_debug_viz gives you a color-coded overlap map (green skin, blue clothes, red accessories, gray residual) that makes tuning painfully obvious.

Outputs

clean_skin, clean_clothes, clean_accessories - the mutually exclusive result - plus residual (pixels that couldn't be confidently claimed) and debug_overlay. The status string reports what happened.

Installing

Standard BrainDead install - ComfyUI Manager search "BrainDead", or clone + pip install -r requirements.txt + restart.

One tuning tip from the pack's own docs: the skin_color_mode defaults are the part people skip and then complain about. If your character has non-standard skin tones and fixed_hsv style results look wrong, confirm adaptive_lab is on - it's the difference between masks that track the actual character and masks that track an average person.