I tested it with generating a man holding a Penrose triangle made of wood. While gpt-image-1 succeeded, nano-banana failed. The aesthetics of nano-banana did lo

I tested it with generating a man holding a Penrose triangle made of wood. While gpt-image-1 succeeded, nano-banana failed. The aesthetics of nano-banana did look much better though. I would guess that it is a diffusion model, based on the fact that it adds irrelevant but pretty background details, which gpt-image-1 tends to avoid.