>>2815823>loras trained on SD 1.5 typically only work with realism models>loras trained on nai or nai-like models typically only work with cartoon/semi-realism models > using regular regularization images assists in the learning process by making sure it's not too rigid with the train directory images (for example, this allows you to put a subject in any environment or any kind of clothing even if every single train image has that person inside and wearing only one type of clothing)So here's my theory. People smarter than me probably have already figured this out but I think this is how it could work. I regularly train realism loras based on real people. Train on SD 1.5. Have a regularization directory of 500 or 600 images depending on who I'm training. That directory is to ensure the model does not overfit. My biggest gripe with training in general is that I can use these models on wheelers and models but they do not work well, if at all, on anime models like aom3, anything v4, Meina, etc.
So here are my questions:
1) what having a regularization directory of let's say $250 realism images and 250 anime / cartoonish images allow it to have more versatility and be able to work with not just SD 1.5 models but anime models too?
2) my theory is that perhaps merging SD 1.5 with another enemy model like Nai or anything V3 could also help with this too.
To the people ITT who are way smarter than me, does what I said and why I wanna accomplish make sense? I want a person-lora to work both with a model like realistic vision AND an orangemix model. Is this possible? Are the methods I described feasible? Would they even work?
