I noticed that almost all generations are white people, so I tried to make a LoRA that gives more diverse results, instant, without prompting it. But it failed, still it can be of some use to get a bit more realistic and less chocolate looking skin.
Also I don't want the 500 buzz to be totally wasted haha, it only works with trigger words black or dark skin
Description
FAQ
Comments (7)
I like your initial thought, but you certainly need to diversify your diversification. We also need;
- latinx
- non-chinese or non-japanese asians
- persian
- ugly, meh and non-model
That was the plan, but since it doesn't take over the skin color at all without just asking for dark skin I don't know how much it will do for the rest.
@JayNL I'm not a pro by any means, but perhaps you can associate the training with a completely random madeup trigger word to prevent associations with the already existing data. From there train multiple ethnicities, skin tones and races under the same trigger word. Setting up an equal number of images for each tone to help randomize the results evenly. Of course this may end up blending (mixed races) but perhaps that can yield results?
Overall, thanks for a much needed attempt.
@Humanity_For_Ai would be nice if I could experiment with that, but retraining costs 500 buzz every time, the LoRA works a lot better on SD3, but that's not running on Civitai...
SHIIIIEEEEEEEET
God bless you!! Its so hard to generate black skin
Bravo, great job on this! Glad someone else spent their buzz making this before I had to. JK Great work!