Remove Background

This is not colour keying. Older background removers looked for a range of similar pixels — the green screen approach — and fell apart the moment the background was busy or the subject shared its colour. What runs here is a segmentation model trained on a very large number of images, which predicts for every pixel how likely it is to belong to the foreground subject.

That is why it copes with a person standing in a normal room rather than in front of a studio backdrop. It has learned what people, products and animals tend to look like, so it can separate them from a background it has never seen before.

It also explains the failure cases. When the model is uncertain — because subject and background are similar in tone, because the edge is genuinely ambiguous, or because the subject is something unusual — the mask gets soft or ragged in exactly those places. It is making an informed guess, not measuring something.

Contrast is what matters most. A subject that differs from its background in brightness or colour separates cleanly; one that blends into it does not. If you control the shot, that single choice will do more than any setting.

Even, diffuse lighting helps too. Hard shadows falling across the boundary between subject and background confuse the edge, and a subject lit from behind often loses its outline entirely. Resolution matters at the margins — a small, heavily compressed image gives the model less to work with, and JPG artefacts around edges are exactly where cut-outs go wrong.

Finally, keep the subject fully in frame where you can. Something cropped by the edge of the picture gives the model no boundary to find on that side, and the result is usually a straight cut along the frame.

The most common use is a product photo for a marketplace listing, where a clean white or transparent background is often a listing requirement rather than a preference. Amazon, eBay and most e-commerce platforms specify it explicitly.

Beyond that: profile pictures and team headshots that need to sit on a brand colour, presentation slides where a floating subject looks far better than a photo with a mismatched rectangle behind it, logos and stickers, and any design work where the subject needs to overlap other elements. Because the output is an ordinary PNG with an alpha channel, every design tool understands it.

Almost everything on this site runs inside your browser. Background removal does not, and the reason is size: the segmentation model is far larger than anything sensible to download into a browser tab, and it needs more compute than a tab can comfortably provide.

So the model runs as a separate process on our machine. Your image is sent over an encrypted HTTPS connection, processed, and both the upload and the result are deleted within an hour. Nothing is retained, indexed or used for training. Pages that upload your files say so rather than implying otherwise.