To demonstrate the system, NVIDIA trained the neural network using the CelebA HQ.
They used an increasingly popular AI training system called a general adversarial network (GAN). The program is fed a massive data set — in this case celebrity photos — and then gets better at creating the desired result (in this case, realistic computer-generated faces) over a period of days or weeks. The "adversarial" component involves pitting two machine-learning programs against one another, with one program "challenging" the other's face creations.
The future applications of rendering realistic looking people — that aren't actually people — seems both potentially useful and unsettling. This is certainly a boon to graphics companies that always need new images of people, perhaps for use in advertising. But it's also important to recognize AI-programs are getting closer and closer to achieving realism in artificial faces; how this might be employed in fake news and other means of deception is unknown — and boundless. For now, though, it's at least given us some really interesting faces to look at.
For its project, NVIDIA found that training the neural network using low-resolution photos of real celebrities and then ramping up to high-res photos helped to both speed up and stabilize the “learning” process, allowing the AI to create “images of unprecedented quality.”
