This one can be a little tricky to work with, so you may have to roll the dice a few times to get something good. It can produce some decent results though. I didn't do a lot of testing with character loras since it involves 2 people and will probably have issues.
It seems to be somewhat picky about the prompts used, but you can grab the prompts from my examples as a starting point.
Training for more steps didn't seem to help things, and if anything the anatomy got worse, so I'm not sure if I'll be able to make it work any better than this.
Description
FAQ
Comments (5)
Work better in str 1.1 - 1.2
Definitely hit or miss but when it hits it's great
This is awesome. I'm really looking forward to a refined version. The future looks amazing for ai.
Why are cunnilingus models ALWAYS woman on woman? If you men aren't doing it you're missing out!
From what I've noticed in training loras, is that it can grab unwanted features from the images used when making concepts. It might take subtle likeness from male anatomy and then gen them onto females, even when tagged male and female, like a more chiseled chin, broader shoulders ect. Just making them less feminine. It doesn't tag facial features and the like (you wouldn't want to cause you don't want same face syndrome). So using all females in the training helps to avoid this.
In regards to using negatives to avoid this the issue with that is the steady rise of popularity of dmd 4 and 8 step models (You can also just run the lora itself on any Xl model). Where you can make an image in much less steps, but doing so does not allow you to use negatives in the prompt... well you can but the dmd will completely ignore it.