You have probably heard at some point in your life, 'Don't believe everything you read!', especially in this day and age of the Internet. That is sage advice, and of particular interest in the era of fake news stories permeating social media. However, with advances in artificial intelligence technology, don't believe everything you see, either. Just check out those mugshots above.
They look real, especially the one on the left. However, those are not actually real people. They are computer generated. A software engineer at Uber created a website that generates new facial images on demand, and rather quickly, whenever the site is refreshed.
"Recently a talented group of researchers at NVIDIA released the current state of the art generative adversarial network, StyleGAN...I have decided to dig into my own pockets and raise some public awareness for this technology. Faces are most salient to our cognition, so I've decided to put that specific pre-trained model up. Their research group has also included pre-trained models for cats, cars, and bedrooms in their repository that you can immediately use," Wang explains.
If you head to the website (click in the link in the Via field below), it will generate a new facial image from scratch from a 512 dimensional vector every time you hit the refresh button in your browser. Some of the results are more realistic than others. In the image at the top, for example, the sunken eyes on the guy on the right hint that it is a fake photo. However, the one on the left is not so easily dissected, and both could pass for actual images.
The underlying technology is rather interesting, and something we wrote about last December. In short, NVIDIA researchers modified a basic Generative Adversarial Network (GAN), which is a relatively new concept in machine learning—it was introduced for the first time in 2014. As explained in a blog post on topic, a common example is a GAN application generating artificial face images by learning from a data set of celebrity images.
NVIDIA's StyleGAN approach improves upon the method by generating artificial images gradually, starting from a low resolution and continuing to a high resolution (1024x1024), and modifying each level separately. This allows the technology to control visual features expressed in each level (face shape, pose, hair color, and so forth) without affecting other levels.
What makes Wang's website so remarkable, though, is how quickly it utilizes this technology. It only takes a few seconds to generate a realistic image. He told Motherboard that his website runs on a rented server with a powerful GPU running Nvidia's software.
"I have it dream up a random face every two seconds, and display that to the world in scalable fashion. Nothing fancy," he said. Wang also told the site that "most people do not understand how good AIs will be at synthesizing images in the future." Duly noted.