NVIDIA inches ever so slightly towards the future, as its new artificial intelligence (AI) created lifelike randomly-generated faces based on existing celebrity faces.
The images are high-quality and on a par with pictures of real people. The researchers of NVIDIA have turned two AI systems against one another in a global adversarial network (GAN) with the goal of producing high-quality photographs of fake people. The two AIs, rather than competing, worked together, with one creating the images and the other refining them in order to produce accurate and intended results.
The two AI systems, called generator and discriminator, also had a library of 30,000 low-resolution images to choose from, most of which are pictures of real-world celebrities. NVIDIA's two AI's achieved the level of and quality needed for photorealism by slowly increasing the pixel count for each image, adding more layers to them.
The GAN network and its two AIs first had to study the process and the goal for 20 days. It is powered by NVIDIA Tesla P100 graphics processing unit (GPU). The same top-of-the-line GPU can also be found in supercomputing data centers.
Still, the AI results could be further refined, as is obvious if their work was to be examined closely. Some of the fake people generated are missing parts of their head or have an unrealistic anatomy. NVIDIA has admitted that while the AI was able to somewhat achieve photorealism, they still have a long way to go before their mistakes get ironed out. As it is now, the algorithm which the AIs use have trouble determining whether shapes should be straight or curves, hence the flawed form of some of the results.
The implications of NVIDIA's AI progress may have numerous significant uses in the near future, but right now, it may take years before actual practical applications for the experiment become viable for public use.
