v1.3 is more artistic and is trained on fuwafuwa mix as base model. It's to most accurate to the character's likeness.
V1.4.1 base model is SD1.5 and V1.4 is nai. Both are more human / anime specific, but fail as a character Lora. V1.4.1 should be used as a cat girl lora.
V1 and V1.2 were earlier experiments trying to tune and experiment with Kohya, base model is Meinamix (my preferred model).
Tags / prompts I found work well were,
Linechu (Trigger Word), a boy with blue hair and cat ears, holding a penguin plushie, wearing a black top, flat chest, slightly chubby,
Keep at strength 1.
Laying on the side is bad for ai prompts, as much as I would like it. various laying poses and non-standing generations always produce weird results. Makes sense if you look at art like olympia, or try drawing trapeze artists.