Background removal is the most requested feature in any listing tool and the one with the worst failure mode. Not because it fails often, but because of how it fails.
The failure nobody notices in time
A cutout model produces a mask: keep this, drop that. When it is confident and wrong, it does not produce an obviously broken image. It produces a clean, plausible, well-lit photo of an item with a piece missing — a pale sleeve read as background, the thin leg of a chair, a glass handle.
You don’t notice because the result looks fine. The buyer notices, because they’re looking at the thing they’re deciding whether to spend money on.
That is worse than no background removal at all. A slightly cluttered photo costs you a bit of polish. A photo of a jacket with one sleeve deleted costs you the sale and, if it ships, the return.
Treating the mask as a claim to check
So the cutout is not trusted on its own. After the mask comes back, a safety pass looks at what it actually did — how much was removed, whether the removal is contiguous with the edges the way real background is, whether the remaining shape is plausibly a whole object.
A mask that fails those checks is rejected, and the original photo is used instead. The result is a pipeline that sometimes declines to remove a background, which is a much cheaper outcome than one that sometimes removes your product.
Why it’s opt-in
Background removal is something you switch on, not something that happens to every photo you upload. Plenty of items — anything on a clean surface, anything where context helps — look better without it.
The general principle runs through the whole app: the automated step produces a proposal, the checks catch the specific way that step fails, and you get the final say. Nothing about “AI-powered” should mean “you find out afterwards.”
