SAMSUNG HAS DONE what Dan Aykroyd and Harold Ramis did 30 years ago with Ghostbusters II and brought a painting to life. The newly-animated Mona Lisa seems a bit more focused on chatting than Vigo was in 1989, but in a world of fake news and deepfakes, it could be nearly as damaging in the long run.
The whole thing is explained along with weirdly soothing music in the video below, or in more depth in the paper here, but here's the crib notes version: Samsung has found a way for AI to generate fake animated videos from just a single frame of a subject.
More frames are better, of course, but the results aren't bad taken from various static pictures. You can see the Mona Lisa merrily yapping away at the 5:08 mark.
The process involves training the algorithm on "landmark" facial features from 7,000 celebrity images harvested from YouTube. The shape of the face, their eyes, mouth shape and other things are dutifully mapped onto a photo, allowing it to move in the same way as the humans the AI has learned from.
It's not just Mona Lisa that the researchers tested it on either: you'll find better results for Albert Einstein and Marilyn Monroe, but obviously, the Mona Lisa is more remarkable as nobody has seen her move in over 500 years.
Yes, it's not perfect yet. There are telltale signs that you're not looking at real footage, other than the fact that, y'know, it's impossible. There's artifacting around the edges of the face if you look closely, but fixing that is probably less of a challenge than getting ol' Lisa moving in the first place, so expect these fakes to get more realistic soon.
That's cool and all, but probably bad news for the world. Certain people are getting adept at dismissing truths as fake news even when there's tangible proof available. How much worse will it be when they can just claim all that proof was generated with a computer and a single photo? µ
Slack, hack and crack
A flaw in the protocol affects iOS, macOS and Windows 10
Wig wearer has issue with non-wig-wearer