What makes gpt-image-2 so good? Is the architecture of the training sets?

it's mostly the architecture, not just the dataset. Reasoning before rendering is the thing that separates it from the pack rn