Remove the background from an image

Most background removers send your photo to their servers, process it there and send the result back. This one works the other way round: the program comes to your device. An image-recognition model is downloaded once into your browser, then it runs on your phone or computer, like an installed app.

The model looks at a small copy of the photo and produces a mask: a greyscale picture that says, for every point, whether it belongs to the subject or to the background. The mask is then refined along the real edges of the photo, enlarged to the original size and applied to your pixels. That is why the final file keeps its full resolution.

The work is done by the device's processor (WebAssembly). The "Portrait" model uses the graphics card when the browser allows it (WebGPU): the first time, its result is compared with the processor's, and the graphics card is only kept if the two agree. Every file the tool needs is served by our own site. No other service is contacted, and no photo leaves your device.

The tool offers three models, because none of them gets everything right. Each one is downloaded only if you choose it.

If the result is not what you expected, switch to another kind of image and run it again: the preview lets you compare within seconds.

• General (42.1 MB): the U²-Net model. It looks for the main element of the picture, whatever it is: a product, an object, an animal. This is the default.
• Portrait (24.7 MB): the MODNet model, trained on photos of people. It keeps hair and full bodies better, but it cannot cut out an object.
• Light (4.36 MB): a reduced version of U²-Net. Handy on a slow connection or a limited data plan, at the cost of less precise outlines. It often keeps the text of a logo better.

The model spots what stands out from the rest. The sharper and better lit the subject, and the more it differs from its background, the cleaner the outline. Some pictures stay hard for any automatic cut-out, whatever the tool.

When you hit a hard case, try the other kind of image, then adjust the outline. If part of the subject is missing, retaking the photo against a simpler background is the better fix.

• Works well: an object on a table or a plain backdrop, a person in front of a wall, an animal on clear ground.
• Logos: on a plain background, "General" keeps the symbol but often drops thin lettering. Try "Light", which holds on to text better.
• Trickier: fine or curly hair, pale fur on a pale background, loose strands.
• Often fails: glass, transparent or shiny objects, smoke, a subject the same colour as its background.
• Ambiguous: several subjects in the picture, or a logo placed over a busy photo. The model keeps whatever catches the eye most, which may not be what you wanted.

In an online shop, on a Facebook page or on a marketplace, a product on a white or plain background is easier to read and makes the whole listing look consistent. Many marketplaces ask for a white background on the main image: check the rules of the site you publish on.

A cut-out does not replace a decent photo. A few simple habits when you shoot save more time than any setting afterwards.

• Put the product on a plain backdrop that contrasts with it: a sheet, a piece of card or a pale wall for a dark object, and the reverse for a pale one.
• Shoot near a window, without flash, and avoid hard shadows that merge with the object.
• Keep the whole product in frame, away from the edges of the picture.
• Then choose the "White" background and the JPG format for a small file, and keep the same framing from one product to the next.

A transparent background lets whatever sits behind the image show through. Choose it if the subject is going onto a poster, a banner, a web page or another design. The file then has to be a PNG or a WebP, the two formats that hold transparency.

A white, coloured or blurred background gives a finished picture, ready to publish as it is. It can be saved as JPG, which opens everywhere and is far smaller than a PNG.

• PNG: no quality loss, transparency kept. The file is heavy for a large photo.
• WebP: transparency kept and a much smaller file. Every current browser reads it, but Safari cannot create it, so the option is switched off there.
• JPG: no transparency. It is only offered with a white, coloured or blurred background.
• In every case the image is re-encoded: the date, camera and GPS position of the original photo are not copied over.

Two sliders act on the mask without running the model again: the preview changes at once, and the full-size file gets exactly the same treatment.

Judge the result on the background you will really use. A fringe you cannot see on the transparency checkerboard may show on white, and the other way round.

• "Adjust the outline" towards "Remove more": clears the haze or fringe of the old background left around the subject. The outline moves slightly inwards.
• "Adjust the outline" towards "Keep more": brings back areas the model left half transparent, such as a pale sleeve or the rim of an object.
• "Soften the outline": blends the edge into the new background. Good for hair and fur, best avoided for a product with crisp edges.
• The colour of the old background is removed automatically from the edge pixels, so no coloured halo appears on the new background.