Neural network specialists Stephan R. Richter, Hassan Abu AlHaija, и Vladlen Koltun created a new way to make GTA 5 graphics to look real.
In some basic projects and old games neural networks were used to upscale the image and to give it more sharpness with creating new pixels. But with new ААА titles that method didn't work well. This new way makes GTA 5 visuals look almost lifelike with special convolutional neural network.
Specialists worked with convolutional neural network and made additional changes to algorithms. It greens the parched grass and hills in GTA's California, adds reflections to the windows and increases the fresnel effect, rebuilds the roads, removes distant haze and so on. The network is using real images from cities to create similar parts on the GTA recordings.
It also using data from the game such as depth, distance to the objects and lighting maps. All these information combined with real footage of the cities creates an almost photorealistic look for GTA 5.
