Background remover

How to remove the background from a photo

1. Drop your photo in

Drag it onto the panel, click to choose, or paste from the clipboard. Nothing is uploaded. The picture is read by this page and stays in the tab.

2. Say what is in the picture

The background uses a model, and is what you want for an ordinary photograph. One colour erases everything within a distance of a colour you click, which is exact on a product shot, a logo or a screenshot. The first run of the model downloads it, once; the colour mode downloads nothing ever.

3. Tidy the edge, then save

Edge moves the boundary in or out. Feather softens it. Then choose what goes behind, nothing, white or a colour, and save at the size you started with.

Why this one does not upload your photo

Because it does not have to. The model is a file, and files can be downloaded to a browser as easily as to a server; once it is here, the work happens on your machine. What crosses the network on your first visit is the model coming to you. Your photograph does not go anywhere, and you can watch that in the Network tab: drop a picture in with the developer tools open and there is no request carrying it. Turn the network off entirely and the tool still works.

The usual arrangement is the other way round. Your picture goes to a server, a model runs there, and a cutout comes back. That works, and it means your photograph is on somebody's machine, subject to their retention policy and their breaches.

The 500-pixel free tier

The well-known background removers are free at 500 pixels on the longest edge. That is a thumbnail. The full-resolution file is a subscription, or credits, and a watermark until you pay.

There is no such tier here, because there is no per-image cost to recover: the computation happens on your machine, not on one we rent. A 6000-pixel photograph comes back 6000 pixels, unwatermarked, on the first go, without an account.

Two models, and how to choose between them

People and portraits is 12.4 MB and is a portrait matting model: it is very good at hair and edges of clothing, fast enough to be pleasant on a phone, and useless on a bicycle. Anything else is 84 MB and general: objects, animals, products, furniture. It is slower and it is the one to use when there is no person in the frame.

Pick the first unless the photo has no person in it. Whichever you load stays cached, so switching back and forth costs the download once each.

The colour mode is not a lesser version of the AI

It is the better tool for a whole class of picture. A product photographed on a white sweep, a logo on a flat field, a screenshot, a scan. In all of those the background is defined by its colour, and a rule about colour is exact where a model is an estimate. It also handles pictures a segmentation model has no idea about, like a diagram or a piece of type.

Click Pick from the image and then the part that should disappear, and raise Tolerance until it is gone. If the subject starts vanishing too, you have gone past the point where the two are distinguishable by colour, and the model is the answer.

Edge and Feather, and the halo they exist for

Every cutout leaves a thin rim of the old background: one or two pixels that were a blend of subject and backdrop and had to be called one or the other. Against a new background that rim reads as a halo, and it is the single thing that makes a cutout look like a cutout.

Edge pulled negative eats it, by moving the boundary inwards. Feather softens the boundary so it sits naturally against a photograph with shallow depth of field. Both act on the mask that has already been computed, so they redraw instantly, and both refine from the original mask each time, so you can move them out and back without the edge degrading.

What it cannot do

It cannot separate a subject from a background that is genuinely the same colour and texture. No tool can, and anything claiming otherwise is guessing on your behalf. It cannot pick one person out of a group; the model finds the subject, not a subject. Glass, smoke and very fine hair against a busy background are where all of these are weakest, and where the colour mode sometimes wins outright.

Which format to save it as

PNG keeps transparency and every pixel, and is the right answer for a cutout you are going to place on something else. WebP also keeps transparency and makes a much smaller file, at the cost of re-encoding. JPG has no transparency at all, if the background is set to nothing, saving as JPG fills it white, and the page says so before you press the button rather than after.