One of the most remarkable elements of Grand Theft Auto V is the closeness of San Andreas real life, but Intel Labs has a recent machine-learning experiment named “Enhancing Photorealism Enhancement” (via Gizmodo). This will help the game look photorealistic.
The processes developed by scientists Stephan R. Richter, Hassan Abu Alhaija, and Vladlen Kolten yield a startling result: a picture appearance that unmistakably resembles the types of pictures that you would casually take from the blurred front window of your car. You have to see it in action to appreciate it very well, but the mixture of lighting, smooth streets, and reckoning vehicles just offers a view of the true road.
The community provides a more detailed and comprehensive description of how the improvement of images really works in their paper (PDF), but we understand that many details have been packed with Cityscapes Dataset – which has been mostly photographed in German Streets. It’s dimmer and from a different point of view, but it almost captures a cleaner, more immersive version of Google Maps’ Street View. It doesn’t act exactly like it’s actual, but it looks like it’s real.
Researchers claim their improvements go beyond other photorealistic conversion processes by also incorporating geometric details from GTA V. As researchers name it, those “G buffers” will contain data such as the distance between items in the game and the camera and the consistency of textures, such as the gloss of the vehicles.
There is no way in which machine learning smarts will expose graphics to higher resolutions, but the graphics of older games were modified by various mod projects on Nvidia Shield TV. In these scenarios, a neural network predicts that a lower resolution game, movie, and TV show will provide the missing pixels of details to achieve the higher resolutions.
This Intel Labs project hopefully shows that there is as much space to improve on the tech side as new consoles and game PCs have on the sheer GPU capacity.
