Trained on one image, training tags were: colored skin, colored sclera, looking at another, 2girls, glowing eyes, profile, portrait, breasts

V1:
The sample images show the lora ran at various weights. I'm still testing combinations of weight/cfg/tags/etc to see what happens, and currently am training another with the same image but different training parameters.

V2e3:
Not necessarily better, just trained different.

I think V1 is the better of the two. With V1 I used autotagging set to .3, and just removed the tags that obviously had nothing to do with the image. It was also trained mostly on default settings. V2 I limited to just 2 tags and used different training parameters (adafactor instead of adam8bit, larger DIM)

I think I'm done experimenting with FLUX for the time being, training is expensive and generating images adds up quick.