Thanks for drawing attention to the whitewashing issue. I always put some work into making my models relatable for a diverse audience, in this case about 10% of the training data is POC woman and white men. In this version, there's not enough data for POC male POV or subjects for it to be promptable. And of course, classic AI issue, the captioning models don't put a white girl tag on there so the concept of "woman" ends up white by default. You should have an easy time prompting with nationality or skin tone tags, but yes unfortunately you'll need "dark skin" or related tags to not get relatively light-skinned examples of any skin tone group.

I'll see about retraining with more images and better tags for brown and dark brown dicks, more Arab and South American women so the dark skin tag doesn't skew toward African heritage as much, and gay POC.

I actually don't see the blonde thing happening though, that might be the base checkpoint. By tag frequency, I see black hair: 195, blonde hair: 127, brown hair: 113. Few with red or colorful hair but it doesn't seem to have trouble prompting those.

Testing CyberRealistic with just the prompt "an american woman with cum on her face" WITHOUT a lora applied generates nothing but women that look a lot like that default. In particular, cum tags seem to push that model toward one look.

It's especially strong on CyberRealistic but it looks like it happens to various degrees on the whole family of "realism merge" type checkpoints that cannibalize each other and contain a lot of the same base content. This lora's dataset included about 400 different-looking women's faces and none of them resemble the sort of default face those models trend to with this lora on, so I don't think it's a direct result of the training data.

Either way, adding weak negative tags related to brown hair, brown eyes, or Asianness seems to balance out the effect and prompting toward other characteristics still seems to work well.

Tried some tags with mostly poor results. Best tries were along the lines of telling it the subject is being washed with soap. "white cum" in the negative or positive seems to be the, uh, cleanest way to adjust translucency. Clear glaze or wet, water, etc. seem to have an effect on translucency but change other things too, and synonyms for cum give slightly different results.

Seems to track with the base checkpoint used, especially CyberRealistic but also similar merges. Easy to get away from with positive prompts, prompting away with negative prompts is hard to do weakly because I think there's no less-than-one weighting in negative but "brown hair, brown eyes" in negative is effective as are various Asian ethnicity tags.

Just added an instructions section, pls enjoy