Interesting. Look up "semantic compression", and you can also try out one of my experiments to see if it works for you.

https://github.com/yuechen-li-dev/GenerativeCompressionProto...