Here is an interesting experiment published by Intel and involving the well-known video game GTA V di Rockstar Games and an intricate algorithm of machine learning able rework polygonal images and transform them into realistic images.

Summarizing everything in plain English: game images and the in-engine information related to the materials, objects, textures and lights used (G-Buffers) are reworked by an image enhancement network which, again in simple terms, uses the frames present in the dataset Cityscape to apply a realistic filter. The image that will be created will then be entrusted, together with other realistic photos to a perceptual discriminator, which will produce the final image. In the words of Intel's engineers, their method should deliver more stable and consistent results when compared to other reprocessing methods such as CUT TSIT.

The operation is particularly complicated, but for more geeks know that the entire documentation is visible within the Intel's GitHub page dedicated to the project.