It's easier to get good results by generating 2-8 images for each prompt, picking the best result, using that as the starting point for a new run.

Tweak the parameters of how close the new image should be to the old one and the prompts, tweak the prompts to address problem areas in the image, generate another 2-8. Rinse and repeat until you have something decent.

I doubt that the really impressive SD images are coming from submitting a prompt once and taking that output. It's better than past open-source efforts like the mini dall-es, but it also still has a way to go before it can reliably produce good results on its own.

AI art on products is an interesting idea, but personally I would use a more fully-featued SD web UI to generate the image locally.