I think gpt-image-2 at least incorporates representations from the base model, even if base model doesn't itself have the output capability. And it does have image input fused directly into it that helps make those representations more usable for image gen, so it's not just generating text.