I trained this LoCon on some pretty bad AI Gen outputs so it's somewhat garbo, but if you don't push it too far some of the outputs are okay. I need to regenerate some cleaner training data and retag, training on AI turns what is 'passable' SD1.5 gens into shit training data.
Some initial flaws with this LoCon that I've noticed- left foot and right foot get mixed up rather easily, though that could very well by the base model. The 'very long hair' tag causes it to flip out. Depth seems to have some issues on some models. Slightly swollen knuckles.
If you're interested, Civitai has both the training data and the LoCon (flawed as it is). The nice thing about using a synthetic dataset is that I can just generate more to fix issues. Assuming that my lack of prompt creativity doesn't do me in ofc💀 https://civitai.com/models/316075/ponyxl-gem-t22
Prompt
<lyco:MIX-GEM-D1 adamw8bit-P1:1> 1girl, animal ear fluff, animal ears, armband, bags under eyes, black shirt, black skirt, blush, breasts, cherry blossoms, cowboy shot, ear blush, extra ears, faux text, fox ears, fox girl, from side, groin, hair between eyes, hand on own chest, heart, heart-shaped pupils, long hair, long sleeves, looking at viewer, midriff, navel, open mouth, petals, pink eyes, pleated skirt, puffy sleeves, raised eyebrows, shirt, side slit, skirt, small breasts, solo, stomach, symbol-shaped pupils, teeth, white hair, BREAK score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up
Negative prompt
(e621_p_low), 3d, sketch, censored, bad anatomy, jpeg artifacts, text
I trained this LoCon on some pretty bad AI Gen outputs so it's somewhat garbo, but if you don't push it too far some of the outputs are okay. I need to regenerate some cleaner training data and retag, training on AI turns what is 'passable' SD1.5 gens into shit training data.
Some initial flaws with this LoCon that I've noticed- left foot and right foot get mixed up rather easily, though that could very well by the base model. The 'very long hair' tag causes it to flip out. Depth seems to have some issues on some models. Slightly swollen knuckles.
If you're interested, Civitai has both the training data and the LoCon (flawed as it is). The nice thing about using a synthetic dataset is that I can just generate more to fix issues. Assuming that my lack of prompt creativity doesn't do me in ofc💀
https://civitai.com/models/316075/ponyxl-gem-t22