I am not an expert when it comes to image generation, but know a fair bit about large language models and machine learning in general. My guess is that like all

I am not an expert when it comes to image generation, but know a fair bit about large language models and machine learning in general. My guess is that like all gradient-based models, they will find certain minima based on the the training data and be somewhat reluctant to deviate too strongly from it as this would most likely penalise both training and test loss. Thus, a "style" can easily "emerge" simple due to the way we formulate these models mathematically. By intuition, consider how say a comedian or workman would often be reluctant to deviate from whatever is tried and tested. Machine learning models tend to be somewhat conservative by default, although there is plenty of literature on how we can sample them to try to avoid this at the cost of reduced accuracy/precision.