I did several trainings almost all experimental as to how to approach what is essentially a reversal of millions of iterations of learning. Example images show good and bad.
The effects on checkpoints seem to have high variability. Juggernaut Aftermath was the checkpoint the test images where generated on. The Lora doesn't appear to effect anime models the same as realistic.
Using the negative African with positive very dark skin gives the most substantial results. Some checkpoints work better then others.
10% of the results look like face swaps,
50% look interesting but not realistic
40% I would consider realistic.
Sometimes the U-Net will simply draw the darker skin as a lighting effect
Note: This training was achieved by using real photos body details from dark skinned females. All images where cropped to only torso, or close up.
Greyscale images of Asian, Middle Eastern, Caucasian ***** female (Head photos) where described accurately and trained at the following rates.
25_Body Details
10_Head Images
This lora could likely be further refined by taking the best AI images and using those in the data set.