Previously, this required much more input.
Researchers from the Samsung Center for the Study of AI Samsung have taught the neural network to create facial animation in a single image. To do this, we developed a new way to overlay personal landmarks: the system can make the “goal” speak, turn its head and apply facial expressions.
So far the result is not perfect, but realistic. This was achieved using a generative, adversary neural network, a type of neural networks consisting of two modules, one of which tries to deceive the other. Thus, the system constantly “learns” and produces a result with a certain level of “realism”.
Similar systems have been before, for example, in 2017, the Deepfake algorithm began to be used to impose celebrities on actors in porn. However, usually such systems required a lot of input data: dozens or even hundreds of images of the “target” face from different angles, angles, with different facial expressions or lighting.
Samsung became one of the first who managed to teach the neural network to create a face in just one frame. Due to this, the researchers managed to “revive” Marilyn Monroe, Albert Einstein, Fyodor Dostoevsky, Mona Lisa, and more.