>take a human and show them 4 or 5 images of an animal they've never seen before they'll generally be able to draw it quite well

4-5 is actually enough to fine tune a pretrained SD model. Which is the correct comparison since we're already pretrained. Even if you ignore all the data upto that point in your life, even newborn brains are pretrained by evolution. They aren't initialised from random weights. Easier to notice this in other animals that can start walking right after birth.