How your image is processed

A short technical account, so the privacy claim on the rest of the site is something you can check rather than something you have to take on faith.

1. The page loads

The pages here are pre-rendered HTML. Opening one downloads text, styles and a small amount of JavaScript. No image-processing code runs and nothing about the tool is initialised yet.

2. You choose an image

The file is read by the browser through the standard file input, drag-and-drop, or a clipboard paste. It is decoded into a canvas in page memory. There is no upload request at this point, and there is no code path in the application that would make one.

3. The model downloads once

The first time a cutout runs, the browser fetches the segmentation model — tens of megabytes of weights — from the asset host of the open-source library the tool uses. That is a plain file download; it carries none of your image. Your browser caches it, so later images skip this step.

4. Segmentation runs on your device

The model executes through WebAssembly on your own processor. It produces a mask: a value per pixel describing how strongly that pixel belongs to the subject. The mask becomes the alpha channel of the result.

When you have drawn a focus area, the region sent to the model is padded outward by about a third on each side, because a salient-object model needs surrounding context to identify a subject at all. The result is then cropped back to your selection.

5. You export

The visible cutout is encoded to PNG, WebP or JPEG by the browser and handed to your downloads folder as a local blob. Nothing is transmitted.

What this costs you

Honesty about the trade-off: local processing means the first run is slow, a low-powered phone is slower than a laptop, and very large images use real memory. A server-side tool would be faster on a weak device. It would also have your photo.

What we do measure

Aggregate counters — page views and tool milestones, in coarse buckets, without cookies, IP storage or fingerprinting. The privacy policy lists them individually, including what the ad network can see, which is a separate matter from the tool.

Where it fails

The model finds salient objects. Photographs with no distinct subject — a sky, a wall, a texture, a dense crowd — produce an empty mask, and the tool reports that as a failure rather than giving you a blank file. The guide on how this works goes into why.