The complexity of putting facial features together and comparing them with one other, as well as human ability to capture the slightest flaws in an image, make the creation of a realistic avatar very difficult.
The program developed by scientists is called Fewshot learning. Using an extensive sampling of celebrity images, the developers taught it to pick facial landmarks from the data base to create the image of a required person.
Fewshot learning consists of three neural networks: embedded, generator and discriminator networks. The first one singles out information that does not depend on a person’s posture from the image; the second one generates new data based on the sampling and the first network’s results; the third one is designed to assess the quality of the image generated. Click the link below to see Fewshot learning in operation.