Blur faces in photos, without uploading them

Blur, pixelate or a solid box?

All three hide a face from someone scrolling past. They differ in how they look and in how much a determined person could still work out.

Blur and pixelate are not proof against recognition. In a 2016 study, researchers trained a neural network on covered photos of 40 known people; it still named the right person for more than 95% of faces pixelated down to about 6 by 7 blocks, and for 58% of faces blurred by YouTube's own face blur, against 2.5% by chance (McPherson and others). That attack needs a list of people to compare against, so it matters most when someone close to the person might go looking.

Our blur first averages each face into a coarse grid and only then smooths it, so the fine detail is not in the saved file to be sharpened back. The cover reaches all the way to the edge of each box or oval and only fades out beyond it. If being recognized by people who know them would put someone at risk, use the solid box.

How well the face finder works

The face finder is YuNet, a small open model from the OpenCV project (MIT license, about 230 KB). Its makers say it finds faces of about 10 to 300 pixels (OpenCV Zoo), so we look at each photo at up to three sizes: the whole photo for large faces, and sharper copies, in tiles, for people further away.

In our test on public domain photos and generated photos of invented people, run through the same code as this page:

Each of these photos took about 1 to 2 seconds on our test laptop, and a photo 4,000 pixels wide about 3 seconds. So the finder is good at faces turned towards the camera, even small ones, and weaker on side profiles, faces hidden by hands or hats, tiny faces at the back and very old or blurry photos. Now and then it boxes something that only looks like a face. Boxes it is less sure about are marked "check" in the list, and faces on posters, screens and in reflections may or may not be found.

What happens to your photos

Everything happens inside this page. The first time you use the tool, your browser downloads the face finder from uncrop.si: the model and the code that runs it, a one-time download of up to 14 MB, which your browser keeps for next time. Your photo is never sent anywhere, and we never see it.

The file you download is a new image drawn from the covered pixels, so whatever was under a box is gone from it, not hidden on a layer. It also leaves out the photo's hidden details (EXIF), such as the GPS location, the camera model and the time it was taken. Your original file is not changed.

The saved copy keeps the width and height of your photo up to 16 megapixels, which covers a 4,000 by 4,000 pixel photo. A bigger photo is covered and saved at 16 megapixels, so the page keeps working on phones, and the tool tells you the new size when it does this. While a photo is still being searched for faces, the download buttons wait, so a copy is never saved before its faces are boxed.

Before you share a photo of other people

Faces are only one way to identify someone. Look the photo over for name badges, school or club logos on uniforms, house numbers, number plates, and screens or papers someone is holding, and drag a box over each one. Check reflections in windows, mirrors and sunglasses too.

Even large services get this wrong now and then: Google blurs faces and number plates in Street View automatically and still offers a Report a problem tool for any face or number plate that needs more blurring (Street View policy). Take the same care with your own photos: open each one and look at the whole picture before you post it.