Most image model comparisons follow the same pattern: put four vendor samples next to each other, pick a winner, and call it a day. That does not help much when you are choosing an API. The annoying questions come later. Can I change one region without touching the rest? How many references can I send? What happens when the text is wrong three times in a row? How much did the image I could actually ship cost? One caveat before getting into it: this is not a blind image-quality test. I did no...