Blur faces in a photo

What this can and cannot promise

Let us be direct about this first, because a privacy tool that oversells itself does real harm. Face detection is very good at faces looking roughly at the camera and progressively worse as they turn away. A profile at the edge of a crowd, a face behind a car window, someone looking down at a phone, a face forty pixels across at the back of a group — any of those can be missed.

So this tool tells you how many faces it covered. That number is the useful part. If you know there were eleven people in the shot and it reports nine, you have learned something immediately. Look at the result before you publish it. Nothing here guarantees anonymity, and any tool that says otherwise is not being straight with you.

Blur, pixelate or block

Blurring looks the most natural and is what most people want. Pixelation is the familiar mosaic and reads unmistakably as deliberate. A solid block removes all doubt and is the right choice when the point is that something has been withheld rather than softened.

There is a real caveat on the first two at low strength. A lightly pixelated or lightly blurred face retains a surprising amount of structure, and there is published research on recovering identities from weak pixelation. That is why the strength control goes as high as it does, and why the default is not gentle. If it matters, use a solid block.

Covering the head, not the face

A detector returns a box around facial features — roughly eyebrows to chin. Hair, ears, jawline and neck sit outside it, and those carry a good deal of recognisable information. The area control expands the covered region beyond the detected box, with extra room above, because hair is up. The default covers the head rather than the face.

The oval shape follows a head more naturally than a rectangle and its edge is feathered, so the result looks like part of the photograph rather than a sticker pasted over it.

How hard to look

Three settings, and the trade-off is real in both directions. Strict only covers unmistakable faces and will miss some. Thorough catches considerably more and will occasionally blur something that is not a face — a patterned cushion, a doorknob, a face in a poster on the wall. For a privacy tool the expensive mistake is the miss, not the blurred doorknob, so thorough is worth reaching for on a crowd.

Metadata is the other half

Blurring the faces does nothing about the GPS coordinates, the camera serial number, and the date and time sitting in the file's EXIF block. A photograph with every face covered can still say exactly where and when it was taken. Run the result through remove image metadata as well — the two together are what "anonymised" actually means.

Up to twenty at once

Drop in a whole set and they come back as a zip, with a total count across all of them. This runs on our server because the detection model does, and each job is deleted as soon as your download has been sent.