Why Your AI Cutout Has a Halo (and the 30-Second Fix)

When a segmentation model decides which pixels belong to the subject, it does it cleanly in the middle of the subject (definitely subject) and cleanly outside (definitely background). The hard part is the boundary, where individual pixels are partly subject and partly background. A blue-shirted person in front of a yellow wall produces edge pixels that are technically a mix of blue and yellow.

When the AI extracts the subject, it has to decide what to do with those mixed pixels. The simple approach keeps them. The result: blue shirt edge plus a faint yellow tint, because the original yellow wall is still partly visible in the boundary pixels. Drop that onto a white background and the yellow halo becomes obvious.

Two reasons. First, your eye is unusually sensitive to color shifts at object edges — it's how we tell whether a cutout is real or composited. A 2-pixel-wide yellow halo is invisible if the photo stays on its original background, because your brain reads it as continuous color. The same halo on a white slide screams 'edited photo' to anyone glancing at it.

Second, modern displays have higher contrast than the displays the original tools were calibrated against five years ago. A halo that was barely visible in 2020 on a typical laptop is glaring on a 2026 OLED or a 4K monitor.

Use a tool that decontaminates the edge during export. BG Clear does this automatically — the export pipeline includes a color-decontamination pass that estimates the original background color at the edge and removes its contribution from the boundary pixels. The result is a transparent PNG where the edge color is the subject's actual color, not a mixture.

If your current tool doesn't do this, the manual fix in Photoshop or Affinity is: extract the cutout, then apply a 1- or 2-pixel inward erode on the alpha channel. This trims the contaminated boundary pixels off entirely. The cutout will look slightly tighter, but it won't have a halo. For most subjects, that trade-off is worth it.

Run a strong test case. Take a photo of a subject in saturated red against a saturated green wall. Cut it out. Drop the result onto pure white. If you see a green tint along the edge — even a faint one — your tool isn't decontaminating. If the edge looks clean, it is.

This is the single biggest free-tool quality differentiator I've seen. Tools that decontaminate produce cutouts you can drop on any background without thinking. Tools that don't make every cutout a small Photoshop session.

If the original background was extremely close in color to the subject, sometimes there's no clean boundary to recover and decontamination just creates new artifacts. For those photos — say, a blonde hair against a sandy beach background — the right move is to reshoot with more contrast or accept that this image needs manual masking.

The other unfixable case is heavy JPEG compression on the source. JPEG compression spreads color across pixel boundaries, which means the contamination is baked into the file before the AI even sees it. A higher-quality source photo solves this. There's no AI magic that recovers what JPEG threw away.

If you have a haloed cutout right now: re-run the photo through a tool that decontaminates (BG Clear is one), drop the new result onto your destination background, and check at 200% zoom. Halo gone in most cases.

If the halo persists, your source photo is the problem. Re-shoot if you can; manually mask if you can't. Don't burn an hour fighting the AI tool — the AI has done its job, and the rest is upstream.