Retrained with a totally brand-new set of dataset.
Added leg tattoo as part of the trigger words.
This version is clearly less overfitting, but with the price of sacrificing some fidelity in some clothing parts (the most noticeable part is her gloves' pattern).
Special training was done on her eyes (grey eyes, blue/cyan pupils), but they may not show the exact colors in certain checkpoints (CetusMix for example).
Her hands will less likely to raise up or appear to be holding stuff (but still they will because those kinds of similiar gestures take up to 60% in the dataset). Unless there're new fan arts (without she holding her weapon), this is probably the best I could do so far without losing much fidelity.
Some other less-important notes:
Why use horns instead of hair ornament (bc they're indeed hair ornament)? Based on testing, under the same training data, the version using horns is easier to undress comparing to hair ornament.
If I added pelvic curtain as part of the trigger words, this version would have been released couple days ago (which I did not). Why? Simply bc I DON'T WANT TO. Humans are lazy, if I could use less trigger words to recreate a charactor, I'll do my best to achieve it (imaging having 10 different trigger words for a single charactor in a Lora, no thanks).
