And if we could to speak the most beautiful characters paintings, or even the pictures of celebrities of these last centuries? Well, in 2019, it's now possible! To give life pictures that, until now, were still ... worthy of a Harry Potter movie? Basically, thanks to artificial intelligence, this technology system is based on a huge database of the model that we want to make speak, but is also based on similar faces, and comes to animate an image in a hyper-realistic way. Come on, we'll explain!
How it works ? The system performs a long meta-learning from a wide choice of videos of the person. Then he is able to select some Plans in particular, that it isolates. Once in stock, he draws a facies resembling in his database fromcustomizable faces"And there comes out a generated version that mix this basic unfamiliar face and video shots of the person you want to talk to.
An approach so well thought that this technology is capable "Learn" of these faces then to create very versions of them realistic and animated: whether from new people or even portraits in painting!
To create a personalized speaking face template, it requires training from a very wide set of pictorial data of one person. However, in many practical scenarios, such personalized talking faces need to be "learned" beforehand, from several pictures of the person, and even potentially, a single image ! Here, these scientists present us a system with an ability that precisely does not need many basic images to make the model talk. And that's what's crazy.
Several recent studies on the subject have shown how realistic these realistic human faces can be achieved by convolutional neural networks to generate them.
For now, according to the study, the only limit of this technology is that it is unable to reproduce facial expressions and other distinctive facial features of people, including looks lively and dynamic. By using other faces as simple bases, when it applies to a celebrity, we immediately realize that something is wrong and that there is a inconsistency. So, let's admit, to create a video fake of someone who speaks, without these inconsistencies, it would be necessary that the system of adaptation is more worked and that the basic face is as close as possible to the person concerned.
The study of this technological system was carried out for the Samsung AI Center and was born at the Moscow Skolkovo Institute of Science and Technology. And for the curious ones among you who wish to know more, the complete study is available to reading by clicking here.
Of course, right away, we could worry about consequences terrible of such technology: to make people say incorrect to people who would never have spoken them in real life. This double impostor created from scratch could cause some future problems ... Nevertheless, reassure us, this system only works for the face and does not yet affect the rest of the body. So, what do you think ?
The Mona Lisa
"The Unknown" by Ivan Kramskoi
Imagined by: Moscow Skolkovo Institute of Science and Technology for the Samsung AI Center