This is the first time I've played with a text-to-image model. I was aware that so-called "prompt engineering" can be tricky, but it's wild to see it for myself. A single space character can be the difference between garbage (or nightmare-fuel) output and output that captures the spirit of the prompt pretty well.
It shouldn't really, have you tried generating a few images with each prompt with and without space?
Even with the same prompt, you can get a wide variety of quality.
Ooh, I've got one of those! I've been getting by trying to run it on my PC, which for various reasons currently has a 5800 and a GTX 1050 4GB, which can just bearly handle optimizedSD at 90s/image but runs out of memory if I try to use the popular webui repo. Swapping to the 5600X might be worth it!
Do you have some examples?
Are you fixing the random seed? If not the variation is more likely to be that than a single space.