Don’t blink: Facebook has a plan to fix your closed-eye photos

We can’t control when we blink, and unfortunately it often coincides with the click of a camera shutter, ruining the picture for everyone else with full mastery of their eyelids. The picture is essentially a write-off, unless you’re adept with Photoshop or prefer to go down the comedy googly eye route.

These pictures may not be a lost cause if clever technology from Facebook ever sees the light of day. The system uses a ‘generative adversarial network’ (GAN) to teach two machine learning systems: the first recognises when eyes are open or closed, and the second does it’s best to fool the first with fakes. It’s a virtuous cycle: the better the second AI gets with its fakes, the better the former has to become at spotting them, leading to further improvements for the second – until eventually human beings will be unable to tell that images have been doctored.

It’s not just picking passable-looking eyes from a database and affixing them Mr Potato Head-style. Instead Facebook uses exemplary data by examining pictures of the same person with their eyes open. From there, GAN can accurately mimic the area around the eyes, from colour and shape to positioning.

You may think sticking fake eyes onto people will look undefinably creepy in some way, but the results are surprisingly impressive, and the research paper is full of examples of ‘before and after’ shots with celebrities suddenly open-eyed. In testing, most human respondents couldn’t say with any confidence which ones were artificially doctored.

The picture on the left is the reference, the middle one is the actual photo, and the one on the right is GAN’s attempt at making a passable fake.

(You may wonder how Facebook found such a rich treasure trove of closed-eye celebrities, and the main source seems to be a Pepsi’s Pakistani Liter of Light campaign, where celebrities from the country posed with their eyes closed with the hashtag #EyesClosedForLight.)

It’s not perfect: colour matching in some photos is a bit iffy, and if the blinking eyes are obscured by glasses or hair, then the results are slightly off. Perhaps more important is the human factor: you may be able to fix poorly-timed blinks, but you can’t sweep away the uneasiness that comes from the knowledge that an algorithm is studying every facet of your face.