After my first attempt at making a LoRA, I took what I learned from that experience, expanded the training set, and re-attempted making the LoRA. Once again, this one was designed to try and create women who appear to be of Watertribe descent as in Avatar: The Last Airbender. I haven't tested if it works for making men of the same descent.

It still seems to have some kinks, but it seems to be working fairly nicely now. I'm not entirely sure how much is the LoRA and how much is the checkpoint, but it's coming out alright; I think I'll try a third attempt when I get the buzz. With some testing, I can recommend the following:

• The LoRA seems to skew cartoony by default, but can be brought to Photorealistic standards by including "Photorealism" in the prompt.

• It may not produce the greatest images by default, but if you include appropriate prompts in the negative field to ward off bad quality (which is just generally good advice with AI image generation), that should fix things. The following is my go to negative for quality control:

  • Ugly, deformed, malformed, animated, anime, painting, b&w.

• Note that the higher the intensity of the LoRA, the more saturated the image will be. I recommend keeping it 0.5-1

The sample images were generated post-production on my own device using Midgard, which was not the LoRA training model.