A presentation of the work of a new type of generative-adversarial neural networks was held at the Samsung Center for Artificial Intelligence in Moscow. What distinguishes it from the previous generation is the addition of a third layer of image processing, which allows any painting to be brought to life. This is a new level of deepfakes - simulation videos with a high degree of realism.
The new AI consists of three subsystems. The "Embedder" neural network studies the source images to determine key points, moving parts and create their displacement vectors. In the case of a portrait of a person, these are lips, eyes, face tilt, cheekbones movement, etc. Then the “Generator” neural network, based on these vectors, generates various versions of the images, which are distinguished from the original by the displacement of individual parts. For example, a palette of emotions and grimaces.
Then the third neural network "Discriminator" comes into operation, which analyzes the final video for defects, inconsistencies and general unreliability. The step-by-step work of the whole complex leads to the creation of fake videos of a high level of realism. The specialization of this AI is complex animation of not just static objects, but those that no one initially expected to see in motion. For example, he is able to "revive" the Mona Lisa from a famous painting.
We have before us a clear example of imitation videos of a new level. It is one thing to make a person in the frame smile or “put on” a hat on a character in a video, that is, to repeat the actions that they could have performed in reality. And another is to depict something that never happened, and most people can only imagine how it should look. However, the new AI not only does this job, but also makes viewers believe in the realism of their fakes.