Photorealistic post-processing of rendered 3D scenes
It is now possible to produce near-photorealistic 3D graphics relatively cheaply; however reaching complete photo-realism requires one order of magnitude more resources. At the same time, deep learning-based super-resolution techniques have proven capable of rendering low-resolution images into photo-realistic higher-resolution images. As such, it may be possible to use related techniques to post-process 3D renders (especially renders featuring people) to become photo-realistic; especially since the remaining issues with 3D renders tend to be at the level of local textures and micro light effects, which a super resolution-type model may be able to handle. Generative adversarial networks may also prove effective (with an adversarial network attempting to disciminate between photos and post-processed 3D renders).
Such a model would stand a chance to eventually end up changing the way a lot of entertainment is produced. If the resulting network can be made to run in real-time, it could serve as a video-game shader, for instance.