Dalle 3 or similar are different models compare to LLM, but in the future OpenAI will probably allow generating image directly from GPT-4o without using the Dalle model (as showed in the demo page), I bet it would much better at following the prompts and generating unexpected things.