Background Remover

This background remover cuts the subject out of a photo and leaves everything behind it transparent, so you can drop the result onto a different background, a product page, a slide or a sticker. It runs entirely inside your browser: the picture is never uploaded, there is no account to create, no watermark on the result and no limit on how many images you process.

How it works

The cutout is produced by U²-Net, a neural network trained for salient object detection — the task of deciding which part of a picture a person would say it is "of". Rather than looking for faces or products specifically, it scores every pixel for how much it belongs to the main subject, which is why it copes with objects it was never explicitly taught.

Your image is scaled to 320×320, normalised, and run through the network, which returns a mask: one number per pixel, from 0 for background to 1 for subject. That mask is scaled back up to your image's real size and written into its alpha channel. The original pixels are never altered — only their transparency — so nothing is resampled or recompressed until you export.

Everything above happens in a Web Worker, a background thread. A cutout takes a second or two of solid computation, and on the main thread that would freeze the page while it ran.

The first-use download

A neural network is a file of learned weights, and it has to be on your device for any of this to work without a server. This one is about 6.7 MB compressed, alongside the WebAssembly runtime that executes it. The tool asks before fetching it, shows the byte count as it arrives, and lets you cancel — a design choice, not a technical necessity. After that it is cached, and the tool works with no connection at all.

The weights are served from this site rather than from a third-party model host. A request to somebody else's CDN would tell them that a particular visitor is removing a background right now, which would quietly undo the privacy the rest of the tool is built for.

Getting a clean edge

The compact model that fits in a web page is good at clear subjects against a distinct background, and much weaker on hair, fur, wire, foliage and anything semi-transparent. Rather than pretend otherwise, the tool gives you the mask to edit:

Edge sharpness pushes the in-between values towards fully transparent or fully solid. Raise it when the cutout has a soft halo; lower it when the edge looks cut out with scissors. Feather blurs the boundary by a pixel or two, which is what stops a hard edge from looking jagged against its new background.

Refine gives you a brush. Erase removes parts the model kept — a shadow, a bit of floor, a second object — and Restore paints back parts it dropped, which is usually a thin strip of hair or a limb against a busy background. Strokes build up, so you can go over an area repeatedly for a stronger effect, and Reset refinements returns to the model's own answer.

Exporting

PNG is the safe choice: lossless, universally supported, and it keeps the alpha channel intact. WebP produces noticeably smaller files with transparency too, and is a better fit for the web. JPEG is deliberately not offered — it cannot store transparency at all, so a cutout saved as JPEG would come back with a solid background and undo the work.