Burger anon, landscape anon, fubumio anon, CDB-anon, style-anon (maybe??), the whole lot-anon, and dark roast anon, I see you guys and love your stuff as always. Busy day though, please forgive me not replying.
Afternoon lads. I decided to test out that dark_fantasy (aka the chocolate skin) LoRA a bit more for the following reasons: 1) its creators intended to do more than make skin dark and 2) they succeeded (good for them but problematic for us choco-chuuba enjoyers).
I ran a test using my Cabalist prompt (since it contains some fantasy element tags like spell-casting and dark energy) with the hope of answering the following questions: 1) what is the best approach if I only want dark skin on my chuuba? and 2) what sort of style will this LoRA give my work if I turn it up?
The image attached to this post is the reference image (no dark skin tags, no dark_fantasy activation tag, no LoRA) and the grid behaves as follows:
https://files.catbox.moe/2yxr0w.png>Top LeftLoRA on (1-0.2 weight), dark skin tags, dark_fantasy tag
>Bottom LeftLoRA on (1-0.2 weight), dark skin tags, no dark_fantasy tag
>Top RightLoRA on (1-0.2 weight), dark skin tags set to zero, dark_fantasy tag
>Bottom RightLoRA on (1-0.2 weight), dark skin tags set to zero, no dark_fantasy tag
TL;DR: to force chocolate girls on a racist model with minimal style bleed, use the LoRA at weight, use the activation tag, use dark skin tags with some emphasis, and turn up your background related tags a bit if you are prompting for that. The last two recommendations will vary depending on your model.
Conclusions: 1) The 'dark_fantasy' activation keyword does not have any appreciable effect on skin color, but it does, in my opinion, make the image look better if you are using the LoRA at weights above 0.2-0.3. It does seem to remove elements of 'realism' from a prompt, so keep that in mind. 2) If you are using a racist model (like mine), you sacrifice some of your own model's quality trying to force dark skin by weighting the relevant tags at 1.5+. The tags are still necessary to get dark skin even with the LoRA, but you can balance them out a bit. 3) the LoRA doesn't like backgrounds, almost certainly trained on an image set consisting primarily of anime/stylized portraits or other subject-focused work with little to no background detail.
