GPT-4o direct image generation

It is already clear that GPT-4o can understand images on its own (and it already does), but the question remains: can it generate images by itself? Or does the generation happen by passing a text prompt to DALL-E? Not so OMNI?