Researchers at Duke University claim to have developed a new algorithm that is so good at upsampling images that it can take a photo that’s not much more than a blurry mess of pixels and turn it into a pretty reasonable facsimile of the original. Nicknamed Pulse (or, Self-Supervised Photo Upsampling via Latent Space Exploration of Generative Models), the system works with photos of people as well, creating realistic faces complete with facial features such as eyelashes and stubble.
However, before you get alarmed at the privacy implications, this algorithm, which uses Generative Adversarial Networks (with two AI systems going head to head. One creates a face, the other tries to determine if it’s real) ‘cannot ‘un-blur’ photos of real people. It only creates a bunch of artificial faces from a low-res image (though some of them may have a rather close resemblance to the person whose photo was used).
While other similar systems exist, Duke claims its method generates images of far greater resolution (64 times the input image) than others. According to report co-author Sachit Menon, this technique has potential in a wide variety of fields – from astronomy and satellite imagery to medical research.
But what makes this new system so much more effective is the approach Duke took: Instead of trying to guess what missing pixels to create based on images it had seen before, their new algorithm works in a different manner – It generates a set of artificial faces and then sees which, when shrunk down to a low resolution, match the input image the best.
And it seems this works – not only does this result in more realistic, error-free images, but when the researchers asked people to compare their results with those obtained using five competing methods of image scaling, Pulse scored the highest, almost matching high-res images of real people!
