Can AI unblur a face? A realistic result can still be wrong

A face goes into an AI tool as a few blurry blocks and comes out with eyes, skin and eyelashes. It looks as though the blur has been undone. But unless you can compare the result with the original person, you don't yet know whether the software recovered their face or made one up.

There is also another possibility: software recognises who is in the picture without producing a clear portrait at all. That is enough to defeat the reason you hid the face.

So there are two questions here. Can we believe the impressive before-and-after? And what should we use on a family photo if blur still leaves clues behind?

The video about undoing pixelation is a useful warning

The video “It's easier than ever to de-censor videos” starts with a pixelated computer window. The text looks hidden when you pause the video, but it moves behind the blocks as the clip plays. Each frame gives away a little more.

Jeff Geerling's accompanying transcript describes how people combined information from those frames to recover the text. They weren't simply making a plausible sentence. The video had retained information about the original.

That changes how to think about covering a video. An unreadable still frame is one view of the edit; someone with the file gets all the others too. Repeated pixelation can leave enough changing detail to work with.

The demonstration concerns text and a particular pixelation method. It doesn't show that any obscured face can be recovered. To understand the face demonstrations, we need to look at what their software was asked to do.

A generated face can be completely imaginary

PULSE, a research project from 2020, searches for a detailed face that becomes a close match to the starting image when shrunk down. Several different faces can satisfy that requirement. Producing one of them doesn't tell you which person was photographed.

The researchers' own explanation explicitly describes the results as imaginary faces. Their paper shows how the method works.

When you watch an “unblur” demonstration, look for the original image and the question the experiment actually tested:

A row of realistic portraits can be a good demonstration of image generation and a poor demonstration of recovering somebody's identity.

A clear face can also come from finding an older, uncovered photograph. That does not show that the blur was reversed. The reverse-image search guide separates finding a copy from matching a face in another picture.

Recognition doesn't need to give you a portrait

In Defeating Image Obfuscation with Deep Learning, researchers trained software using examples of known people, then asked it to recognise other photos of those people after blur or pixelation had been applied. It succeeded with several of the tested methods. The task was to match people from a known set, not identify any stranger on earth. Read the original research.

This is why looking at your edit and thinking “I can't recognise that face” doesn't settle the question. Software may use patterns you don't notice, and a person who already knows the family may need much less: the account name, a school uniform, the child's voice or a caption saying who they are.

The practical decision is therefore about what remains in the whole post, not just whether the face looks blurry enough at phone size.

Make the file reveal less

Blur and mosaic still draw a version of the face. A solid cover replaces the pixels inside it with a colour. If the aim is to remove those facial pixels from the file you share, choose Solid, crop the person out, or leave the picture unpublished.

The cover has to include the whole face and stay with it throughout the clip. A solid circle over the eyes leaves the mouth and jaw visible; a cover that stops before the video does leaves uncovered frames. Our comparison of blur, mosaic and covers explains those differences, and the video guide shows how to adjust a cover that drifts.

Save a new photo or video with the edit built into it. Then open that saved file and watch it with the sound on. Pause at turns, entrances and quick movement. Check reflected faces as well as the ones you deliberately covered, and listen for names or places you meant to keep private.

When you post, select that edited file. The untouched original may still be beside it in your photo library. Check the caption and audience before sending it too; neither is changed by a face cover.

Stacking effects isn't a security proof

Our earlier mosaic used several passes. A test of that implementation found that extra passes did little about a soft edge that was letting original detail show through. Changing the coverage addressed that leak.

The current tool and the recorded tests are explained in how our face covering works.

There's no need to upload your child's original to an unfamiliar “unblur” site to check whether your edit worked. That would give another service the face you're trying to withhold. Choose a method that removes the facial pixels, make sure it covers the right area, and check the file you actually send.