Trained from Hel... Dishwasher1910's work, based on PonyXL v6.
V03, cleaned up dataset.
Something relevant/ irrelevant:
I've been testing it for over a day now, and this is the second run of training.
For the first run, I thought I did something wrong since the style doesn't follow the anime style strictly, I thought "hmm, maybe it is trained for too short". This run I doubled the steps, still the same... When it's showing Dishwasher's style, it is really good, but sometime it'll just gen something reaaaaaaaly western cartoonish.
I asked around, realize that's how it is, the art style is just going everywhere, let me know if I did something wrong, either for training or prompting. I've tried the neg prompts on other 2d anime style gens, it helps, but still, occasionally there'll be a LOL style output😂
I've tried to use different "score_x_up", it changes shading/details, but the higher the score is the more bias toward western cartoon style it shows, and lower score is more forgiving for anime style based on my limited knowledge of the model.
Or, maybe my lora is built too small, maybe a big dimension would help, but I won't know, this is the biggest size I can make with the GPU I have ヾ(•ω•`)o without spending a whole week to make a lora.
So... have fun, it's a hobby, not a job, right?