Google’s neural networks now boast the ability to recover damaged or heavily pixelated images — something that we have grown accustomed seeing only in products of popular culture such as the TV series CSI.

According to the company, the new Artificial Intelligence (AI) in question can enhance an eight-pixel square image, increase its resolution by 16 fold, and effectively restore lost data.

The neural network, as stated by the company, could play a key role in allowing users to increase the resolution of blurred or pixelated faces in a way that was previously thought practically impossible. A demo was also presented wherein the company enhanced several images of bedrooms from 8×8 to 32×32 pixel.

According to Google researchers associated [PDF] with the project, the neural network gathers the extra information by sort of “hallucinating”. Apparently, the system was given an extensive training that included seeing and analyzing innumerable images of faces so it can learn typical facial features.

Meanwhile, a second part of the system compares the 8×8 pixel images with various 32×32 pixel images that they could possibly be shrunken versions of. Both networks then work in unison to combine and redraw the best guesses of what the original facial features could be like. The AI effectively enforces a significant improvement over previous methods based on up-sampling.

Worth noting, the system is, of course, prone to occasional mistakes as it can only make educated guesses based on the knowledge of how facial features generally are. Just in case it doesn’t have enough info successfully redraw a face, the output could be unidentifiable with the same person as the original (heavily pixelated) image. Odds are also that sometimes it will end up drawing totally funny (or weird) looking objects instead of a face.

However, with all factors taken into account, the system is said to be capable enough to pull off a decent job around 90% of the times, says the team behind the project.