My very first LoRa. Settling with this result since the more I train and experiments, the result got worst. Also, with my outdated PC each training takes 4+ hours to finish, might be better to release it into the wild to fish for tips before experimenting more. 95% created from in-game screenshots trained against the awesome NED model.
Trigger words
nophica, 1woman, redhair, green eyes, green dress, leaf circlet = should work generally
pink pupil, leather choker, armlet = can be added but might not changing much
leaf thigh strap = can be used to prompt result above knees
I might have over prompt on these, thinking I had to be as detailed as I can over the datasets. Let me know if this is actually unnecessary
Recommendation
Clip Skip 2
HiRes fix at least 1.5 (Examples use Denoise 0.5 and 4k-Ultrasharp)
The TI's in the negative prompts used in the examples especially bad-hands-5.
Weight 0.6 ~ 0.8. I recommend 0.7
I use motherly, milf, curvy to get the momma look
Links:
v0.5
An accidental training I did that somehow able to provide an eye appealing realistic pictures. But, in my opinion, it is more to applying 'nophica like' style to the character.
I don't even remember the training parameter for it and unable to replicate it.
Strongly recommend using 0.9 for the weight.
Tips?
Appreciate any tips/guide to get the model more tame/adaptable. Not satisfied yet with the lighting hard to tone it down and small characteristic that the AI can't get it right.