Looking at the picture below, you may think it’s just a picture in real life. In fact, this is a “picture enhanced” gta5 game screenshot, does it look very realistic? The feeling is almost the same as the picture in reality.
Compared with the early “paper player” game, 3A has changed dramatically. Not only is the modeling quality far better than before, but also thanks to the continuous development of image quality engine and hardware technology. Now in many games, we can see the game scenes with exquisite image quality and gorgeous special effects.
However, in some sandbox games that focus on real-world simulation, even if the picture is more beautiful, we can still distinguish it from the picture in the game at a glance. In the words of “low EQ”, the texture of the game picture is still not close to the real world.
Recently, researchers at Intel Labs are thinking about how to use machine learning to make rendered game images look closer to the real world. If this technology can be implemented and put into wide use, it may improve the game image quality to a new level.
In the summary of this Intel project, how to enhance the synthetic image by convolutional neural network (a deep learning algorithm) is described. The team will train the neural network with the information from the game’s rendered image and the middle rendering buffer (g buffer); This intermediate rendering buffer provides the information about the geometry, material and illumination in the scene, and is used as the input of image enhancement network to modulate image features. Finally, through the image enhancement network, a visual image similar to real life is formed.
In fact, as like as two peas in our life, we want to draw a copy of the gourd. We must understand the appearance and size of the gourd and other parameters. These parameters are in your mind. When you paint the gourd, you will use the parameters stored in the brain. In this way, the gourd can have a higher similarity.
However, there are still some problems with this technology. For example, the license plate becomes blurred, the color is paler compared with the rest of the video, and there are still some differences between the color in reality.
In a word, Intel’s project has a certain progressive significance for the development of the game field, especially the game image quality, which means that the game field can make full use of deep learning technology to improve the game experience. At present, the widely used technologies include DLSS technology of supersampling deep learning and RTX ray tracing technology of simulating real ray rendering, If this project of Intel in the future can be carried out smoothly and successfully, there may be an unprecedented visual subversion in the field of games or vr virtual reality in the future.