Remove the Background from a Picture

A product shot for a listing, a portrait for a document, a picture for a slide — all of them need the background gone. Here it is done by a neural network that runs inside this tab, so the photo never leaves your computer.

or drop them here

A photo of a person, an animal or an object. JPG, PNG, WebP, HEIC, up to 30 MB

The picture stays on this device: it is not uploaded anywhere. The model is downloaded once and stays in the browser — after that the tool works with the network off.

The model here is a light one, four megabytes. It handles a clear subject on a distinct background well, and struggles with fine hair, fur and transparent edges — paid services run models forty times heavier on their own servers.

Nearby: Photo Editor · Crop Image · Passport Photo Sheet · Image Converter

How to use

  1. Choose a picture — a person, an animal or an object photographed as the main subject.
  2. Press “Remove the background”. The first run downloads the model, about seven megabytes; after that it is already in the browser.
  3. Leave the background transparent or put a plain colour behind, then download the PNG.

Good to know

The cut-out is made here, not on a server

The model — a light network called u2netp, four megabytes — is downloaded once and runs inside this tab. That is why the photo is not uploaded: there is nowhere for it to go. The core that executes the model is shared with the neural voices on this site, so anyone who has used those downloads only the model itself. After the first run the tool works with the network switched off entirely.

What a light model does well, and what it does not

A clear subject against a background that differs from it — a bottle on a table, a person against a wall, a dog on grass — comes out cleanly. Fine hair, fur and transparent edges are where it gives up: the outline gets a soft fringe instead of separate strands. Paid services run models forty times heavier on their own machines, and they do that part better. It is a fair trade, but only if it is said out loud rather than promised away.

When nothing worked, the page says so

A neural network can fail in three ways, and each one looks like a working answer if you only glance at it. It can find no subject at all and return an empty cut-out. It can take the whole frame for the subject and change nothing. Or it can be unsure everywhere, leaving a grey fringe along the whole outline. The mask is checked against all three before the result appears, and you are told which one happened — rather than handed a file that looks finished.

Frequently asked questions

Is my photo sent anywhere?

No. The model is downloaded to your browser and runs there; the picture is read from the disk into the page and never goes back out. You can check it yourself: after the first run, switch the network off — the tool keeps working.

Why does the first run download about seven megabytes?

That is the model itself (four megabytes) plus the core that executes it (under three). The core is shared with the text-to-speech tool here, so if you have used that, only the model is downloaded. Both stay in the browser afterwards.

Can I put a white or coloured background instead of a transparent one?

Yes. Pick “a plain colour” under the result and choose the shade you need — white for a marketplace listing, for example. Transparent is the default because it is the one you cannot get back later.

Why are the edges of the hair messy?

Because separating individual strands is the hardest part of the task, and the model here is deliberately small so that it fits into a browser download. On hair, fur and lace it draws a soft outline instead of single strands. A shot against a plain wall comes out noticeably better.

Does it work on a phone?

Yes, at about the same speed — a second or two per picture in our measurements, slightly slower on a phone than on a laptop. The download is the same seven megabytes, so on a metered connection it is worth doing once over Wi-Fi.

Related tools