By slowly increasing variation in pixel patterns, recognised from a massive database of real people, NVIDIA’s A.I can shift and morph these patterns into believable, real faces. All up until the point that near perfection is found in around five days. When similar programs can be used to recognise images of animals or objects like trees and vehicles, it’s fun to theorise how massively this will impact on our videogames. Check out the video below and bear in mind – none of the faces after “Generated Images” are real.
As you can see from the video, NVIDIA believed it would help the A.I’s learning curve stability, to start off with a very low resolution set of pixels. Gradually, that grid of pixels is increased into higher resolutions until more massive variations in colour per pixel could be refined accross the board. Until, eventually – voila. You have a face of a person who doesn’t exist, created by two neural networks. One that takes in information from CelebA HQ, and another that evaluates and morphs it. Objects like cars or airplanes are evidently easier for the A.I to generate, while the complexity of face generation still has a way to go. Nonetheless – this is some trippy stuff.
This mind boggling example of face creation from NVIDIA is the result of just one system – CelebA HQ. Bear in mind the potential for variety in other future systems that don’t yet exist or systems that could be combined. Then their results could even be at higher resolutions in the future as computing power continues to grow. Game developers could literally talk to their computer and say something along the lines of “I want my character to have a round face, with red hair and blue eyes”. That developer could then wait for a day or two and evaluate his hundreds of options. The concept of graphics, in the dim and distant future, could well end up a thing of the past.
NVIDIA’s researchers had this to say about their latest pursuit of photorealism “While the quality of our results is generally high compared to earlier work, and the training is stable in large resolutions, there is a long way to true photorealism. That said, we feel that convincing realism may now be within reach.”