Grand Theft Auto V is one of the better-looking games in the modern console generation(s), but calling it photorealistic would be a bit of a stretch. However, a scientific paper made use of GTA V to test out a convolutional network that greatly improves the game's graphics to more realistic levels.
In a paper titled 'Enhancing Photorealism Enhancement,' Stephan Richter, Hassan Abu AlHaija, and Vladlen Koltun talk about a new technology that can improve efforts to "enhance the realism of synthetic images." To do this, they trained a convolutional network to recreate images from Grand Theft Auto V frame by frame to the Cityscapes Dataset at an "interactive" rate. The images in Cityscapes are generally from German cities and recorded with an average car camera. However, by using Cityscapes as a baseline for how reality looks like, the network is able to change the graphics in Grand Theft Auto V to make car paint glossier, pave the asphalt roads, and make California's sun-dried plantlife more vibrant and green.
While there are photorealism technologies out there, the video above shows that they have plenty of artifacting and distortions. To avoid this, this technology makes use of, among other things, G-buffers from the game's engine and a modified sampling strategy to help stabilize the images and keep them looking stable for longer. Perhaps most interestingly the technology is not being done in a post-render or anything, but while the game is running.
The paper and explanation as to how this technology works are pretty hefty, but it's an impressive look at what may be the next big step in graphics technology.
Grand Theft Auto V is available on many different consoles and PC via multiple storefronts, including Steam. As for the paper, you can find the abstract here, along with links to the full paper, videos, and comparisons.
