An attempt at training a style LoRA based on the 300+ illustrations made for the Wakfu MMORPG, this includes all the half body artwork made for the NPCs' dialogue interface, plus all the 18 playable class artworks for both male and female characters.Since different artists have taken part in making the illustrations over the years (the game is about 15 years old at this point), it is a mix of all of the styles, but i think the end result gets the intended job done personally.The keyworkd is: "WakfuMMOArt"Please note that since the training data is almost entirely composed of white/transparent background images of an upper body focus, the LoRA works best when you emphasize the character and not the background and do a little closeup, at least seems to be, share your results if you wish, this is my one attempt at understanding style LoRAs. Pony V2 was trained with tag shuffling and that's pretty much the only difference between V1, the dataset for training was the same on all 3 versions.The races of the characters are also included in the training tags to see if they would help output specific ones, but in my experiences it doesn't seem to do much. Iop and Cra were probably too short of a tag so they are "wakfuiop" and "wakfucra" to prevent mixing unexpected results.Rogue is "Roublard", Masqueraider is "Zobal", Sadidas are separated into "Male sadida" and "Female sadida" because the males all have a bush of hair/beard covering their face, and also tried to do that with Osamodas since the males have more variety in their designs.The rest of the class names are their english names preceded by male and female or on their own and a space between the two words. Tell me how it goes, this is mostly a shot in the dark to see what works, and it gets the job done for the most part.
