2026-07-25

Why background removal gets hair wrong, and how to clean it up

Run an automatic background remover on a portrait and the outline around a shirt or a phone usually comes out clean. Run the same tool on loose hair, and the edge often turns rough — a staircase of tiny blocky steps instead of individual strands. That's not a bug in one particular tool. It's the predictable result of two genuinely different problems, and most background-removal tools, NearBG included, solve one of them and not the other. Here's what each problem actually is, why hair specifically exposes the gap, what NearBG's model does about it, and what its touch-up brush can and can't fix afterward.

Segmentation and matting are different problems

"Remove the background" sounds like one task. Under the hood it's usually one of two different ones. Segmentation asks a yes-or-no question about every pixel: is this foreground or background? The answer is a mask that's effectively binary — each pixel is in or out. Matting asks a different question: what fraction of this pixel belongs to the foreground? The answer is a continuous number between 0 and 1, not a label. That's not a minor implementation detail. It's a different kind of output, built for a different situation.

Why hair specifically breaks a binary decision

A single strand of hair is often thinner than a pixel at normal photo resolution. Where it crosses in front of a background, the pixel underneath isn't purely hair or purely background — it's a genuine blend of both colors. A segmentation model still has to force that blended pixel into one bucket. Whichever way it goes, part of that pixel's true color is wrong. One pixel like that is invisible. A hairline made of thousands of them, each forced into a binary choice, adds up to a visibly blocky edge instead of the soft, semi-transparent fringe real hair actually has. The same thing happens at fur, lace, and the fuzzy edge of a dandelion — anywhere the true boundary isn't a clean line.

What NearBG's model does about it

NearBG removes backgrounds with U2Netp, a segmentation model, chosen specifically because it loads and runs reliably inside every visitor's browser without a server (a separate tradeoff, covered in this blog's first post). Being a segmentation model, not a matting model, is part of that same package: it outputs a mask close to the binary kind described above, not a soft alpha gradient. Fine, hair-strand-level matting is explicitly not something the current model attempts. That's a disclosed limitation of the shipped tool, not an unnoticed gap — and it's the direct, visible cost of the same "runs everywhere" tradeoff the model choice made for a different reason.

What the touch-up brush actually fixes

After automatic removal finishes, the result sits on an editable canvas, and a brush lets you clean it up by hand — no upload, the same zero-network rule as everything else in the tool. The brush only erases: it clears alpha under the cursor and leaves the rest of the image untouched. That makes it good for specific, real problems — a patch of background the model mistakenly kept, a stray fragment detached from the subject, a messy patch along an otherwise clean silhouette. It does not add pixels back. If a few hair strands were already merged into the background and dropped from the mask, no amount of brushing recovers them, because there's nothing left to reveal — the brush can only subtract, never restore partial detail that was never captured. A reset button reverts to the untouched automatic result if a cleanup attempt goes wrong, and brush size is adjustable for finer or coarser passes.

Reading a photo before you upload it

Because the failure mode is predictable, so is roughly when it will show up. A subject with sharp, solid edges against a plain, high-contrast background — a product shot on white, a face against a dark wall — segments cleanly, because there's little ambiguous blended-color territory to force a binary decision on. A subject with fine wispy hair, low contrast against its background (dark hair on a wood-paneled wall, for example), motion blur, or transparent or reflective material like glasses lenses or sheer fabric, gives the model far more ambiguous pixels to guess wrong on. None of that means the tool won't work on those photos — it means expect visible edge roughness there, and plan on the touch-up brush for cleanup rather than treating a rough result as something gone wrong.