Isn't image generation basically doing "anti semantic ablation", starting with a blank canvas and iteratively refining it into a meaningful collection of pixels

Isn't image generation basically doing "anti semantic ablation", starting with a blank canvas and iteratively refining it into a meaningful collection of pixels? Is it possible to do the same thing with word generation, such that it sharpens into an opinionated version (even if it would do something different each time?)