AI learned to recreate a human portrait from a set of pixels

The legendary tool of the cinematic special services, which allows you to get a detailed photo from a blurry picture, has just become incredibly close to reality. Researchers from Duke University (USA) have developed the PULSE algorithm, which creates a photorealistic portrait from a highly pixelated picture. This is done by a special artificial intelligence.

This AI is a kind of adversarial generative neural networks, trained to work with photographs of people. One neural network uses a pixel image as a set of starting conditions and generates a pseudo-photo, the second evaluates the result and issues edits.

The algorithm is impressive in its capabilities - its authors promise a 64-fold increase in image detail. A picture in a resolution of 16x16 pixels, from which it is only clear that this is a semblance of a face, turns into an image with a resolution of 1024x1024 pixels. It can be easily confused with a real photo, and the neural network even takes care of such trifles as the pattern on the collar, if appropriate.

Due to the fact that the program does not recreate the original image, but only generates suitable copies, it will be difficult to use it for serious tasks. On the other hand, the algorithm can be optimized to recreate the appearance of distant exoplanets, not for scientific, but for artistic purposes. Or to visualize the microcosm, to show bacteria and viruses in such a way that they become understandable and interesting not only for specialists, but also for schoolchildren.