Remove Background from Photo

How the Photo Background Remover works

Remove background from photo: what to know first

Every "remove background" website you have used sends your photo to a server, runs a model there, and sends back the result — which is why they need an account, a watermark or a per-image fee, and why you should think twice before uploading a passport photo to one. The same job can be done in the browser. This page loads MediaPipe's selfie segmenter, a 250 KB neural network that estimates for each pixel how likely it is to be part of a person, and uses that estimate as the alpha channel of a new PNG.

What the model is and is not good at. It was trained on people — portraits, selfies, video-call framing — so a head and shoulders, a full-length figure or a group all cut out well. It is not a general object segmenter: a cat, a product, a car or a logo will produce nonsense, and the colour-key transparent PNG tool is the right choice for those if they sit on a flat background. It also works at a modest internal resolution, so on a 4000-pixel photo the edge is smooth but not hair-strand precise; for a profile picture, a slide, a CV or a shop listing it is more than enough.

The two sliders. Edge threshold is the probability at which a pixel switches from background to person. At 50 the model's own opinion is used; drag lower when it has cut off an ear, a hand or the top of the hair, higher when a slice of chair or wall has come along for the ride. Edge softness sets the width of the ramp between fully transparent and fully opaque. Zero gives a hard edge that looks cut with scissors; ten to twenty gives an anti-aliased edge that sits naturally on any background; higher values start to look like a vignette and are mostly useful for deliberately soft portraits.

Transparent versus Colour. Transparent saves a PNG with an alpha channel, which is what you want for anything that will be placed over something else — a slide, a web page, a composite. Colour flattens the cut-out onto a solid fill and is the shortcut for ID and passport photos, where most guidelines ask for plain white or off-white, or for a product shot on a brand colour. The colour is applied at save time, so you can preview both and only download the one you need.

Resolution and file size. The output is the same pixel size as the input; nothing is downscaled. A 12-megapixel phone photo produces a 12-megapixel PNG, which with a transparent background is often 10–20 MB because PNG compresses photographs poorly. If the picture is destined for the web, run it through the image resizer after, not before — the model gets a better edge from the full-size original.

Privacy. The photo is read into memory by your browser, the model runs in WebAssembly inside the page, and the PNG is generated with the canvas API and downloaded. No request carries any image data, which you can confirm in the browser's network panel. The model and runtime themselves are about 12 MB, fetched once and cached.

Related tools