I continue to try to train LoRA locally. This LoRA generates images of girls in a Megumin costume. Sometimes it gives out poorly drawn hands, sometimes pixelated faces. I managed to select prompts for high-quality generation, you can see the examples.
I achieved the best results with Base SDXL and Ratatoskr models, but RealVis sometimes not bad too.
I used str 1, cfg 6.0, dpmpp_2m_sde, karras