BD Remove Background

Background removal is the most commoditized operation in this ecosystem - BiRefNet is in core ComfyUI now, rembg has been around since 2020 - so the question isn't "can it remove the background" anymore, it's "can it survive hair." BD Remove Background is the BrainDead answer: a self-contained, text-grounded cutter that chains SAM3 segmentation to pymatting edge refinement, and it's the one I'd reach for when the subject has flyaway hair or a complex edge and a plain matte just won't do.

What it is

One node that runs the whole pipeline internally: SAM3 text-grounded segmentation → hole fill → optional pymatting alpha matting → RGBA output. You don't wire a SAM3 model, a rmbg node, and a matting node together - it loads everything itself and gives you a clean cutout in one shot.

How it works

You give it an image and one or more positive prompts (prompts, one per line, union-combined - typically person, character, product, foreground object). It runs SAM3 for each and unions the masks. Negative prompts (negative_prompts, default background) get unioned and subtracted, which is the trick for punching the backdrop back out of a mask that bled into it.

Then the real work: matting_mode. The default closed_form runs pymatting's closed-form alpha matting around the mask boundary to produce soft, hair-accurate transparency instead of the hard chunky edge SAM3 gives you. none skips it if you want speed and hard edges. The trimap band is controlled by matting_erode / matting_dilate (8 px each by default - wider band = more matting area, slower). Around that you've got a whole toolbox: fill_holes_radius seals interior gaps, decontaminate removes background color spill at edges (that faint white fringe you see on black backgrounds), edge_refine runs a guided filter or VitMatte (auto-downloads hustvl/vitmatte on first use) on the boundary ROI, edge_shrink trims a stubborn halo, and key_gaps punches background-colored islands out of gaps between legs and fingers.

The inputs that matter for beginners

• prompts - the subject, one per line. This decides what gets kept.
• matting_mode - closed_form for soft edges, none for fast/hard.
• external_mask + mask_mode - hand it a mask from any source (ATR parser, another SAM3 node, manual) and it either constrains SAM3's result, subtracts, or skips SAM3 entirely and uses your mask as the alpha.
• crop_to_content - crops the output to the mask bounding box, killing excess transparent border.
• sticker_outline - a fun one: adds a die-cut colored trim around the subject, like a sticker.

Outputs: rgba (subject + alpha), mask, rgb_white_bg and rgb_black_bg composites (white is the one you want for 3D pipelines), crop_box, and sticker. The white-background composite is worth calling out - that's what you feed an image-to-3D node like Pixal3D or TRELLIS when you need a clean subject on a neutral field.

Gotchas

One that'll bite you: ComfyUI's LoadImage MASK output is the inverse of the subject (subject = black, background = white), so invert_external_mask defaults to ON to compensate. If you wire a mask straight out of SAM3 (which is already subject=white) you need to turn that off or everything inverts. Second: closed_form matting needs pymatting installed - it's not in the pack's requirements.txt, so pip install pymatting if the node complains. VitMatte downloads itself on first use.

Installing it

ComfyUI Manager: search "BrainDead" → install. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/BizaNator/ComfyUI-BrainDead
cd ComfyUI-BrainDead
pip install -r requirements.txt
pip install pymatting   # if you want closed_form matting

Restart, find it under 🧠BrainDead/Segmentation. The SAM3 checkpoint (Comfy-Org/sam3.1) auto-downloads on first run, so the first cutout takes a while - subsequent ones are fast. If hair still looks like a mowed lawn after matting, raise matting_dilate a touch and let decontaminate eat the fringe; that combination fixes most complaints.