Now that it's available, I decided to throw the only non-joke natural language dataset I have at ZIT training and see what happens. Unfortunately training failed at 6/10ths of the way through with no sample pictures except for one epoch. Might as well post what it did make.<s> </s><s>I have no idea</s><s> if this is even a functional LoRA or a broken corrupt file</s> (looks like it functions, actually much better than the Flux version, but lacking). I expect I'll unpublish this and post a proper version as its own page when ZIT training is not totally broken. (edit: In retrospect, no reason I can't just make it a V2. Infact this broken one works better than what I got fully trained)Trigger: A Golden Sun style illustration of [character description], “Adventurer” and “elemental spirit” are used frequent. “The background is plain white, making the character the focal point.” was used for all white background characters, while the Djinn are the same but use “creature” instead of “character”Made with 32 images, adding the missing box art illustration I forgot in Flux.Fully Trained Version: I don't think this actually works properly. Looks weird.

Files with SHA256: 878636ebf275d9467bc493e4d5f9804467eaea7b5134a7c51773f03be10c55c5