Researchers at Nvidia have presented an algorithm (or rather a GAN - Generative Adversarial Networks) that can synthetically generate facial images across all ages and races. The morphing sequences look particularly impressive in the accompanying video, when one face is almost imperceptibly transformed into another and one face is "recombined" from another.
What is special about the new method is the possibility of influencing certain aspects of the result - although neural networks could previously be trained with the input of many looking faces to generate new faces, this was done in a black box, i.e. without the possibility of influencing individual elements such as the hairstyle.
With the help of the new method, however, individual aspects can be specifically changed, such as gender, skin colour, hairstyle, age or facial expression. The quality of the generated faces is surprisingly good compared to many predecessors - they look very realistic.
Although the algorithm can only create static two-dimensional faces, in combination with other techniques such as this one it would also be conceivable to use artificially created realistic faces in moving sequences.
Recombination of faces from sources A and B
Funnily enough, due to its universality as a neural network, the algorithm also works with other objects such as cars or beds, which can also be regenerated at will, as the Democlip shows at the end. The algorithm has to be trained with photos of these objects instead of faces. Here the detailed research paper.