After researching why some gens looks strange is because stable tries to combine lightning from digital photography inserted in 2D illustrations giving sometimes results 2.5Dish and some other realistic details. Better put in negatives like realistic, photo, detailed lightning, tags that are from photography and see if takes better gens. May or could not may work at all since all checkpoint base models and extra loras could fluctuate results.
Discussion
Using the word realistic is always bad in prompting because realistic cannot be measured by the bots judgement. It actually causes the bot to generate more unrealistic pictures because by definition realistic is close to but not quite so it's trying to be close to but not like a real photo.
You might want to avoid "natural" in prompting as well. An interesting thing I read while back in promt writing was that you get better results when you put in the specs of a digital camera example.
Blue skyline over (Manhattan, New York), (photographed with Cannon Rebel EOS T5)+, (Canon EF 100-300mm F4.5-5.6 USM Lens), (ISO 1800, -0.2 exposure)+++
Look up some camera models, focuses lenses, & that cameras settings feats & see what pops out of the prompts. Send images my way when you do!
I just use local, so bot judgement is not an issue lol and for the rest like I say depends on checkpoint but not going to saturate negatives beyond 200 since xl doesn't need a lot schizo prompt.
Localhosting is the only way it should be ran. Who trust a online service 99.9% have ip embed on so your ip addresses is in the pixels of the generated image.