I've had nothing but grief with these AI online models, and my only conclusion is that they ignore their own prompt guides. Consistently get extra, missing, deformed, mutilated, amputee, etc... even though prompting in negatives to not have them.
The first 10 pictures were made by just repeating the same step over and over again, ignoring the entire prompt guide, and it produced more consistent good results than following the prompt guide, so honestly not sure if I can produce better pictures with online options.
It was a nice experiment, but until I have a rig powerful enough to do it locally, with my own modifiable model or multiple LoRas, I don't see myself being able to produce anything in this thread with online options. Oh well. Its cool that these kinds of stuff is possible with local production, but its a shame that online versions don't even come close.