This is the 2nd version of the model.
It overal keep its strenghts, but it now generate a consistent character as intended.
"Monochrome" and "grayscale" are affected by the lora, and they will "often" produce images with colored eyes and/or hairs (change the strenght to fix of those tags or use the "BREAK" keyword to control the output more consistently).
The new versions all work better with colored images, and they all are good in postprocessing in the in-paint section (i particularly like the output of the in-painting on face at 0.65 strenght without changing anything on the prompt and affecting only selection.
There are 5 versions:
Soft, it's the version that less impact on the style of the images, so it is the actual recomended one if you plan to mix styles.
Mid, it's a version in the midpoint between soft and full. It impacts the style, but not too much (among those three versions is the one that required less, in-painting on my generations)
Hard, it's fine and reliable.
Full, It's a version that impact on the style of the image and that generate often "sharp" faces, like intended.
Complete, If you can't choose between the 3 previous ones, you can pick thise one. It's pretty versatile, but it requires you to play with the lora strenght.
N.B: I in-painted only faces to increanse some details, but the results were pretty ok even without doing it.
I suggest a strenght between 0.4-0.9 (it is the suggested value for each of those models), around 0.7 the results are almost always the one intended.
This version has only 1 trigger word:
Lynn
It is not required to add: "1girl, solo", but it recomended to achieve better results.
There are still keywords that are affected by the lora:
monochrome, grayscale, holding, hair bow, bow, sword, ponytail, ribbon, hair ribbon, portrait, smile, open mouth, upper torso
In case of issues with one of those tags you could fix it by changing the strenght of the keyword or placing a "BREAK" between the keyword and the lora in the prompt. (I noted great differences in the output using the last mentioned method, with the same seeds)
(I did also some test with clipskip and it looks like it works also in clipskip 3, but ofc it's not recomended)
