[danschuller]

I think you could do style transfer on a framebuffer before pushing it out and that might be quite interesting, especially if you can have it look correct over time.

The larger possibility is that video has shown the AI has some capacity to preserve structure overtime. The objects in the video maintain their dimensions and appearance as we would expect. This suggests there's some representation of objects or an equivalent to it. If that could be controlled more then potentially you have the ability to hold something like a game in working memory and iterate on parts of it.