Okay Novowels, i give up, here are characters.
Another experiment model based on re:zero characters. For now only 12 characters are more or less reliably usable. Others didn’t have enough solo art, or a didn’t have time to gather their character + character folders yet. That applies for most of side characters.
I had no time to gather more 2 characters for now, and no folders with 3 and more characters yet, so you’ll have a hard time infering more than 2 characters at a time without using additional tools. 2 character inference is supported, but only for limited character set, like Ram + Rem. Other than that, you’ll have hard time trying to get together other characters, other than main cast.
While quality is quite mid by my standards(for characters), i don’t have hardware capability to make better at the current moment in time, as multi-character models are harder to train. If you want to help with that – https://ko-fi.com/anzhc xD
So far characters that are supported: Echidna, Rem, Ram, Elsa, Frederica, Felix, Subaru, Priscilla, Otto, Emilia, Beatrice, Petra
Beatrice and Petra are trained only on their SFW fan-art.
I plan to expand it as much as i can, though, my patience and sleeping time is already at limits, as training took over 10 hours at current dataset already.
Im just testing how things go on Locon model with wider arrangement of characters when you don’t balance them out. In this particular dataset i tried to keep character datasets at 100 images, but some, like Echidna and Rem are based on 170 and 125 respectively. Less popular characters, like Frederica and Otto, don’t have much images, so their datasets are scarse. That is also a reason why current version is limited only to 12 characters, while total folder amount of unique characters is ~16-18, some of them, like Patrasche, almost don’t have art, and solo image was only one in that particular case on Danbooru. This will be offset with combo folders of dataset in future, like curently i have Ram + Rem, or Rem + Emilia and so on, but that requires time, and especially training time.
Current version was trained with 10 classification images per dataset one to not train style(to respect artists’ work in that regard) and train character features better, which amounted to over 11k images that i had to generate for roughly 2 hours straight, to just start training, which totalled at over 110k steps. I will drop usage of class for future trainings, unless by miracle someone gifts me a server dedicated to train stuff, so i can save time and sanity. I will not be concerned with copying anyones’ style at dataset counts of over 2k images, but that is also going to be over 100k steps of training, which will mean ~8-12 hours of training.
I do strip artists names, so it is not possible to infer their style without model that already has them embedded.