It still blows my mind that midjourney & co. can get decent illumination, shadows and even reflections. Just... how?!
Wisdom of the crowd?
It's just statistically inferring from previously seen images.
I’m not sure that sufficiently explains an apparent ability to do three dimensional reasoning when performing statistical inference in two dimensions?
Apparently millions of shadows and billions of parameters is enough.
Does it always get them right, or just most of the time?
I've definitely seen it get reflections/lighting/shadows wrong. In exactly that sort of "this is an unusual perspective and it can't actually do the math" way.
Artists today are still getting them wrong. Art education puts emphasis on this stuff today, but people make mistakes.