The neural network creates realistic human faces from scratch

Artificial intelligence is now able to create realistic images of human faces that are difficult to distinguish from real ones.

Researchers at NVIDIA have been working for years on an algorithm that can generate realistic-looking people, and now the Generative Adversarial Networks (GAN) neural network has created “photos” that are very difficult to distinguish from real people. People created by artificial intelligence are formed using a method known as “style transfer”. The neural network mixes the characteristics of one image with another. The generator considers the image as a combination of three styles: “coarse styles” (posture, hair, face shape), “medium styles” (facial features and eyes) and “subtle styles” (color scheme).

The developers report that the algorithm is also capable of taking pictures of small animals or objects. But the system is still not perfect, and there are signs that can produce a fake image. For example, the most difficult part to recreate is hair. They are difficult to form realistically, so they often look colored or unnatural on the images of a neural network.

naukatv.ru