That is a good question. My guess is that their training data was a mix of photorealistic images and historical paintings, and thus the outcome is the cartoony

That is a good question. My guess is that their training data was a mix of photorealistic images and historical paintings, and thus the outcome is the cartoony style that looks somewhat realistic but not overly so. There’s also the possibility that they avoid being too realistic for fear of deepfake scares WRT the underlying tech. And so if you want images that look realistic without being believably “real”, you’ll get something like the default aesthetic of Midjourney.