>>52289458Wrong image
>>52289147I remember using 4x-AnimeSharp and it keeps screwing hand into object for no reason, I change to 4x-UltraSharp,
where animsharp derived from, but it doesnt f*ck with it anymore. I don't know why but it did the job.
>>52289540I actually want to gen her with chubby to test versatility, but idk what the tag is lol
>>52289716Lora training size are determined by the dimension network (not resolution image) they were trained with, it affect size the more higher dimension within scale of 2^n. Higher dim learn more details from the dataset, but at cost of putting more list (which called dictionary in numpy format) to the tensor.
It's not necessarily higher dim, ez training, for character lora, you only need 32 dimension since most of detail are on upper body, for style lora 128 dim would be enough to diverse more detail for the model. Higher than that probably used for learning terrain or background
or some object that has too much detail on it 