Remove background from image

What makes this AI background remover different?

Most background removers send your photo to a server. This one runs the AI on your own device, so the photo is never uploaded, and the result is free at full size.

It picks its model for your device. A computer with WebGPU gets BiRefNet, a larger model that keeps posters, logos and cars solid. A phone, or a browser without WebGPU, gets ormbg, a lighter one.

And we show our working. Both models are open and named, their licences are linked under the tool, and the tables below give our own test results, including where each model fails.

How does it work without uploading your photo?

The cutting is done by an AI model that runs inside this page, in your browser. Your photo is read from your device, cut there, and handed back as a download. It is never sent anywhere.

The model looks at the photo at 1024 by 1024 pixels and decides, pixel by pixel, how much of each belongs to the subject.

That map is then stretched back over your photo at its own size, so the cut-out keeps every pixel of the original. A photo wider or taller than 4096 pixels is scaled to 4096 first, which keeps it inside the canvas an iPhone allows.

Three more steps follow the model. If the subject is small in the frame, the tool looks a second time at just that part, so a small subject gets as many of the model's pixels as a large one.

Then it cleans the edge. A pixel that was half subject and half old background gets the subject's own colour, so no fringe follows it onto a new background. On the lighter model it also clears faint haze left far from the subject.

There are two models. A computer whose browser supports WebGPU runs BiRefNet, an open model under the MIT License, on its graphics card.

A phone or tablet with WebGPU runs ormbg, an open model under the Apache License 2.0, on its graphics card. Without WebGPU, or with Data Saver on, ormbg runs on the processor.

On photos of people the two give much the same cut-out. On a poster, a logo or a glossy car, BiRefNet keeps the subject solid where ormbg leaves parts of it see-through. That is why a computer's graphics card gets the larger model, and a phone, with less memory, the lighter one.

Both files come from this site, not from a third party, and the requests for them carry nothing about your photo. Your browser keeps them, so the next photo, today or next week, starts straight away.

The tool says which size your browser will get before anything downloads. On a phone, use Wi-Fi for that first photo, or turn on Data Saver to take the smaller download.

How does the bulk background remover work?

Choose or drop several photos at once and the page works as a bulk background remover. The photos queue up and are cut one after another by the same AI model, and each result appears as it finishes.

One at a time is deliberate. The model can only work on one photo at once, and a phone that opened thirty photos together would run out of memory. The model downloads once for the whole batch.

Pick Transparent, White or a colour for every photo. You can change it after the run without cutting again, because the colour is only added when you download.

Rename any photo in the field above it, or name the whole batch as a base and a number, such as shop-01 and shop-02. Then download everything as one ZIP, or one photo at a time.

There is no limit on the number of photos, only time. Here is a batch of 100 through this page:

The memory stays flat because each finished cut-out is written to your browser's private storage on your device, not held in the page. In a private window, where that storage is missing, the tool pauses and asks you to download what is done.

If the browser crashes or the tab closes mid-batch, open this page again. It offers the finished cut-outs back; photos it had not reached need adding again.

The touch-up brush works on one photo at a time. To fix a cut-out by hand, add that photo on its own.

How long does it take?

On the Mac we build on, a cut takes about five seconds with WebGPU on the larger model and about four on the processor alone, which runs the smaller model.

A small subject takes about twice as long, because the tool looks twice. You see its first cut after the first look, about halfway, and the sharper one replaces it when it is ready. A big photo takes little longer than a small one, because both models always work at 1024 by 1024.

An emulated phone with WebGPU, which gets the lighter model, cut a 1200 by 675 photo in one second on the same Mac. We have not timed it on a real phone yet.

What does it cut out well, and what does it miss?

Both models do well on people: hair, glasses and shoulders against a busy street or a plain wall. They part ways on pictures that are not a portrait.

We ran three harder pictures through both. These are the results:

Before choosing ormbg for the processor, we tried it on thirteen photos against a smaller model other in-browser tools use:

MODNet's 8-bit version, the small download, was worse again: it wrecked the old print, the cat and the shoe. Our 8-bit ormbg looked the same as its 16-bit version, at half the download.

Both models are weaker where the subject and the background are the same colour, and on anything see-through, like a glass or a veil. Check the edges in the preview, and fix what is wrong with the Touch up tools before you download.

On a phone, or without WebGPU, graphics are weaker too. Flat areas of colour can come out partly see-through, so a colour wheel looks faded.

On a computer with WebGPU, a picture with no single subject can lose part of itself. The model keeps what it takes to be the subject, so an advert full of text may lose some of the text and keep scraps of the rest. Paint what you want back with Keep under Touch up, and tap a scrap with Remove piece to take it out whole.

How does it compare with remove.bg and Adobe?

We read each service's own pages on 2 October 2026. This is what they state, beside what this tool does:

Here is what that means for a 12-megapixel phone photo. This tool hands it back at 12 megapixels, free. remove.bg's free result is 0.25 megapixels, and the full size needs a paid plan.

They are ahead in two places. remove.bg goes up to 50 megapixels on a paid plan, past our 4096-pixel limit. And a server asks nothing of your device, so on a phone or an old laptop it may well be faster.

This tool is ahead where it matters for most photos: the free result is full size, there is no account, and the photo stays with you.

Can you use it for a passport or visa photo?

Often not, and the reason is the rules, not the tool. Several passport and visa offices ask for a photo that has not been edited, and some name background removal outright.

For those, stand against a plain light wall and take the photo again. For a profile picture, a product listing or a school project, a cut-out is exactly what you want.

Transparent PNG, or a colour behind it?

Choose Transparent and the download is a PNG, the format that keeps the see-through parts. Put it over any design, slide or page and only the subject shows.

Choose White or a colour and the download is a JPG with that colour baked in. Nothing is transparent any more, so a JPG is the better file.

The difference is large. Our 1920 by 1440 photo of a dog came out as a 3.05 MB PNG, and as a 519 KB JPG on white.

Is your photo private?

Yes. The photo never leaves your device: it is opened, cut and saved by this page in your browser. There is no account, no server copy and no watermark.

What does travel is the other way: the model and its engine, downloaded once from this site. Those requests carry no photo data.

Common questions