Tomoe Carmine (Production kawaii) PonyXL LoRA -
https://mega.nz/folder/dfBQDDBJ#3RLMrU3gZmO6uj167o-YZg/file/BC4iUIYQExtremely limited dataset. 29 model shots and some art.
Mako Sameshima (PRISM...) PonyXL LoRA -
https://mega.nz/folder/dfBQDDBJ#3RLMrU3gZmO6uj167o-YZg/file/YawCSSRbReduced dataset to mostly coherent fan art and model screenshots, seemed to do well with the bumped up learning rate. Outfit is less accurate in exchange for significantly more flexibility. Face seems to stay accurate 90%+ of the time, hair ornament seems better than ever
Alias Anono (V4Mirai) PonyXL LoRA -
https://mega.nz/folder/dfBQDDBJ#3RLMrU3gZmO6uj167o-YZg/file/dC4UySxSExtremely limited dataset. It seems the autotagger didn't get the hood, or maybe it was tagged as a headdress, I haven't checked yet. The reason why I haven't made sure is because it already gets the eyes wrong half the time so I want to figure that out before anything else. Anyone have any luck with heterochromia on SDXL? I have another config with a higher learning rate that seemed to work very well with omocat and Mako but I'm wondering if there's anything else I can do. Would removing all eye color tags help? In my brief experience with removing tags or some things not being tagged, it has pretty much just resulted in the unet doing whatever it wants even with an activation token that would supposedly inherit untagged details.
I am considering renting an A40 for a couple hours to quickly run through at least the character loras I have queued up. In total I have 97 loras I'm at least mildly interested in porting/finishing for pony
