I made this LoRA to test how I can create one with just few images.

I started only with 22 images, some of which are quite low quality. I trained a LoRA on these images, generated 15 good images, painted over weird errors and then put them together to train a new LoRA on the 22+15=37 images.

Looking at the tests on four different models, epoch 14 seems to be the most consistent, so that is the one I upload here.

Activation tag is just Roberta. No need to specify blue skin or black hair, but adding furry and animal ears seem to help.