Huh, I wonder if they trained it by feeding it architectural renders and "what actually got built" photos...

It's probably just prompt based. Actual fine-tuning for these kind of use cases is getting less common than it used to be.

Likely. You can go into Nano Banana or ChatGPT right now, upload a pretty architectural rendering, and tell it to make it look old, weathered, winter, etc and it will come out looking very similar. Give it an example to really dial it in.