@john Have you compared the results to your own prompting attempts?
I have 2 conflicting arguments.
LLM created prompts could use more streamlined wording making it easier for the image generator to hook. Which results in better generated pictures.
Putting stuff through more GenAI filters out human "entropy",which is substantial for good generated content. Resulting in worse, less interesting results. Like changing money back an forth which results in less money because of fees.