Finding the faces

Free forever. No sign-up, no watermark, no limits — ads keep it that way.

Step two. Your photo has already been decoded and is sitting in this browser tab. The job now is to find the faces in it, which your own device does, using a small detection model it fetches once.

What is happening while the bar moves

The face detector is a file: Google's MediaPipe BlazeFace short-range model, a few hundred kilobytes of it. The progress bar is that file arriving — a real percentage of a real download, not a spinner pretending to be busy.

It is fetched now, after you chose a photo, and not when the page first loaded. That is deliberate: making every visitor pull down a model before they had even decided to use the tool would spend their data and slow the page for nothing. Once it has arrived your browser caches it the ordinary way, so the second photo you redact today skips this wait.

Your photo is not being sent anywhere

This is the moment people most often stop and wonder, so here is the plain answer. Nothing is uploaded. The traffic runs one way only — the model comes to your image; your image does not go to the model. There is no server here that receives photographs, which is firmer than a promise not to look: there is nothing to look at and nowhere for it to sit. The page on how it works traces the whole path a file takes.

If it finds nothing

"No faces found" is a normal outcome, not an error, and the tool says so rather than failing. Two causes account for most of it. A head seen fully side-on falls outside what a short-range face detector is built to recognise. And a face that is small in the frame — someone forty metres back in a street scene — may be a few dozen pixels across, which is not enough to go on.

Either way the answer is the same: draw the box yourself on the next step. Dragging a rectangle over a head takes a couple of seconds, and it is exactly how the license plate tool works, since there is no automatic plate detection on this site at all.

This model looks for faces and nothing else. It does not look for number plates, house numbers, or text of any kind. No dependable, permissively licensed in-browser model exists for those, so they are covered by hand with a box, the same way you would black out text in a document photo.

If it finds a face that is not there

Detectors sometimes mark a face in a poster on the wall, a portrait in a frame, a reflection in a window, or a pattern that happens to sit face-shaped in the grain of something. Nothing is broken when that happens. Select that box on the next step and delete it. A false positive costs you one tap; a face it failed to mark costs rather more, so the detector is not tuned to be shy.

Why the boxes look too big

Each box is padded outwards from the face the model actually located. A tight crop around eyes, nose and mouth leaves the hairline, the ears, the jaw and the chin showing, and for anyone who already knows the person that is plenty to go on. The margin is there so the box covers a head rather than a face. Any box that takes in more of the picture than you want can be pulled back in on the next step.

When the bar finishes, the faces appear as boxes over your photo. Carry on to adjusting the boxes, where you also pick how each one is covered: a solid bar replaces the pixels underneath and cannot be undone, while pixelation and blur can sometimes be partly reversed.