Researches at Intel Labs used machine learning to recreate and amalgamate real-life street scenes, which drastically enhance scenes from the popular video game, GTA V, into frighteningly realistic graphics.
The Grand Theft Auto V game has the reputation of having the most attention to detail ever to be developed into an open-world video game. However, as amazing, intense and creative GTA V is, the game is by no means photo-realistic.
What the reachers at Intel Labs have done is quite literally modify the images from the game with the help of tens of thousands of dashcam sample images. Then, they make use of machine learning to take the rendered game images and filter them through their in-house Image Enhancement Network to produced their enhanced image results.
The Image Enhancement Network extracts the G-buffers (3D geometric buffers which are responsible for generating scenes that sell the illusion of highlights, reflections, distance to and from the camera etc.) from the original game footage. Then, the complex algorithms match up these broken down scenes from the original game with that of real-life footage to generate the incredible outcome.
Where Intel is now with this project is having a small yet playable section of GTA V that can sample the enhanced scenes in real-time to give the most realistic graphics ever in GTA V.
Take a look at the video below by the YouTube channel, Intel ISL, on Enhancing Photorealism Enhancement...