Lunala LoRA trained on NovelAI with Kohya training scripts using the DAdaptAdam optimizer at 512x512 resolution. Very spicy, I recommend adding (dark purple skin, wings) to your prompt to improve consistency. Skin color and patterns are inconsistent. Wing pattern and shape can be unstable. Recommended to lower the LoRA strength if you're running into issues. Despite the problems, good results are possible. Should have wide compatibility with model mixes based on NovelAI, which is most of them. Trained on an all female dataset.