Intel is currently working on a new machine learning project as part of their Intel Labs program called Enhancing Photorealism Enhancement. They have taken none other than GTA V as their flagship test dummy to implement some of the rendering methods available as part of their “photorealism” project.
In short, the machine learning technique takes real-world photos and combines them with the rendering output for GTA V. This results in some rather fantastic-looking visuals that look next to real life. Part of the realism comes from the washed-out lighting and incredibly reflective cars that drive across the camera.
Intel says by taking real-world cityscape images from an asset pack made up of German towns and cities, they were able to automatically implement tweaks to GTA V. This starts off by adding reflective maps to the cars, smoothing out the road asphalt, and crazily enough, automatically adds greenery to the mountains in the game.
The rendering pipeline seems complicated at first but Intel found a way to carefully refine the workflow in order to create an image with almost no distortion and some extremely enhanced photorealism. Intel says they first take the rendered image in GTA V, pass it through an Image Enhanced Network while extracting the G-buffers from the game engine. G-buffers contain the geometric information for what the game renders including the depth of 3D models, textures and lighting.
Intel then passes all this G-buffer information into the Image Enhancement Network. This is then encoded into a Perceptual Discriminator alongside the real-world photos. All of this then combines together to create an automatic, machine-created render that drastically improves the quality of the game. Keep in mind that this rendering method is not simply adding a texture pack to the game. Instead, it combines real-world photos to automatically attach a list of new items and lighting to the game. Pretty incredible.
You can watch the full breakdown of the Intel Enhancing Photorealism Enhancement pipeline down below. It is great to watch and see how Intel achieved the image.