Researchers Terro Karras, Samuli Laine and Timo Aila of NVIDIA have developed a generative adversarial network (GAN) AI that creates realistic human face portraits that don’t originate in real life. Instead, the generator views certain features, such as hair or freckles, are used as data points to help the AI intuit the rest.
We propose an alternative generator architecture for generative adversarial networks, borrowing from style transfer literature. The new architecture leads to an automatically learned, unsupervised separation of high-level attributes (e.g., pose and identity when trained on human faces) and stochastic variation in the generated images (e.g., freckles, hair), and it enables intuitive, scale-specific control of the synthesis.
This all sounds scary, right? Well, coding artist Kevin McDonald has posted a very helpful article that shows how to determine whether a photo is real or AI developed. He points to asymmetry, weird teeth, messy hair and a strange watercolor appearance.
At low resolutions, almost all the images in the paper are indistinguishable from photographs. There are only a few artifacts that stand out to me that I will try to address.