I went ahead and hosted the app to test (https://gemini-tryon.trieve.ai/). It's solid! One note is that the social icons appear to have incorrect href's.
Performance wise, Gemini is so damn close to being able to do this reliably. I didn't test, but I would expect that OpenAI's image gen setup would be slightly better.
Yes, one thing that I noticed was that grunt of the work was in the prompt; the more descriptive the prompt, the better. Also sometimes it generates a completely different image.