The models are deliberately engineered to create “good” images, just like cameras get autofocus, anti-shake and stuff. There are many tools that will auto-prettify people, not so many for the reverse.

There are enough imperfect images around for the model to know what that looks like.