There is this guy who tests every new image generation model by asking it for a pelican riding a bicycle. I found some of the results interesting, because they

There is this guy who tests every new image generation model by asking it for a pelican riding a bicycle. I found some of the results interesting, because they broke down the subjects to their basic shapes and reordered them geometrically instead of physically. It was a new perspective on how to think about a bicycle, and perhaps the AI's most authentic voice. And I'm sure the AI was punished for it and trained towards more photorealism, because pastiche is supposed to be the default. But that, too, is a conscious choice made by the trainers and prompters.