Remove a Photo Background Without Uploading It: How On-Device AI Works

What actually happens when a remover asks you to upload

The standard deal goes like this: you hand a website your photo, its servers run a model, and you get a cutout back. The cutout is usually fine. What you also handed over is the photo itself — plus everything inside the file. Most phone photos carry EXIF metadata, and that often includes the GPS coordinates of where the picture was taken. Upload the file and the location goes with it.

How long the copy stays on their server, who can access it and whether it trains future models is a matter of that site's privacy policy, which most of us accept in about four seconds. For a product shot, nobody cares. For family photos, a client's unreleased product, or anything with an ID document in frame, the calculation changes.

How on-device removal works

The alternative flips the order. Instead of sending your photo to the model, the model comes to your photo. Modern browsers can run neural networks locally — through WebGPU where it is available, or WebAssembly everywhere else. The page downloads a segmentation model once, your photo never leaves the tab, and the result is composited on your own hardware.

Some real numbers, from the tool linked below. On desktop it loads a full segmentation model of about 91 MB — a noticeable first download, then cached, so repeat visits start fast. Processing a typical photo takes a few seconds on a recent laptop. Phones are different: mobile browsers cap how much memory a tab may use, well below what the big model needs, so on phones (and on desktop browsers that fall back) a lighter model of roughly 4.6 MB does the work. It is quick and still useful — but its edges, especially around hair, are visibly softer. An honest tool tells you which model ran. Ours does.

One more step matters. A photo often contains several separate subjects — two people, a person and a dog. After the preview, you click the subject you want to keep instead of accepting whatever the model guessed. Picking part of one subject (just the left shoe, say) is not something segmentation can do reliably, and any tool claiming otherwise is overselling.

Step by step

1. Open the Background Remover and choose a photo. It loads into the page — nothing is sent anywhere.
2. Check the preview. What you see is the actual image the model will work on.
3. Click "Remove the background" and wait for the on-device model to finish.
4. If the photo has several separate subjects, click the one to keep.
5. Download the transparent PNG.

Where on-device struggles — honestly

Flyaway hair, glass, veils and smoke are hard for every segmentation model, server or not; expect to see the difference on those edges. Very large photos — a 48-megapixel phone shot — can strain a phone browser's memory limits. Old hardware runs the WebAssembly path slowly. And that first model download on a slow connection is genuinely annoying. None of this is hidden: it is the price of keeping the file home.

When uploading is still the right choice

If you process hundreds of product photos a day, need an API, or remove backgrounds from video, a server service earns its fee. Batch pipelines and integrations are what servers are for. The case for on-device is the everyday one: a handful of personal photos, on a machine you own, where "we deleted it afterwards, probably" is not quite the guarantee you wanted.

You can check the claim yourself, by the way. Open your browser's developer tools, watch the Network tab while you remove a background below: after the model files load, processing produces no upload traffic at all. That is a stronger promise than any privacy policy paragraph — because you verified it, not us.