>>66487729Sorry, I forgot the link
https://rentry.org/CCC_TrainingThe settings are meant for character loras that get baked easily due to bad dataset so you will probably have more lenience to increase some of the numbers.
I highly recommend you go through the document you linked before, first and read it once.
>>66493260NTA. I'm pretty sure the limit in SDXL models is not the billions of tokens they understand but the poor understanding and conceptualization of logic between the clip models and the unet, together with the lacking refinement of the model on visually appealing illustrations and anime images.
A model will load up 12gb of vram sitting idle regardless if it has 6 milion ponies in it or 0. The end result is just a 0.7263637437 becoming 0.745236432 when you mention the word pony. The number was there before and it remains there after the training, it's just changed. I bet your SD1.5 model already knows what a pony is, it will just draw a brown mini-horse. As for capturing as much information as possible, just look at hll. it has 100 chuubas and 1000 styles fit in an 800mb lora, meanwhile a character lora is 32 mb so it's a waste of space.
If you want to have fun, just download a bunch of Uma Musume characters, find out what are the pony tags they used and overwrite them with a lora with Uma musume girls. Job done, you overwrite all of the pony training with anime girls.
I'm not on either side of things but I just want to see where things go.
As for resource limiting factors, Team Red is waiting for you. AMD just announced rx7600xt with 16gb of vram and the X600xt series is usually a $250 card.