Textures usually ship compressed in a format that is supported by the hardware. Hardware texture compression saves a lot of bandwidth and therefore improves runtime performance. That compression step is done offline and it's slow enough not to do it on the client's side, otherwise just using ordinary JPEG compression would reduce file sizes substantially.
Shipping size is usually not the limiting factor for texture size, GPU memory is. Texture compression must work on small blocks for random pixel addressing.
Which doesn't entirely rule out a deep learning based compressor, just makes it much harder :)
Why use repeated textures? Just provide a low resolution "pattern definition", say some blades of grass, and fresh unique textures could be generated by a network. probably pretty computationally cheap too
Fwiw, I have seen a NN texture compressor that is compatible with game engines and doesn’t require any AI data to decompress. I can’t give any details because it’s not my project and not published yet. I just wanted to say I know for a fact that your idea there is a good one. ;)
Train the game upscaling AI with high resolution textures, then ship the game with lower res textures. High quality rendering with smaller file sizes, though won't be as good as the original pictures.
Not very useful in reality, but would be a pretty neat proof of concept.
The question is: would the AI data be smaller than shipping the high quality textures in the first place?