This is the first successful LoRA training attemp I have had on Dreambooth and I unfortunately was not able to train it on SDXL but I'll try training an SDXL version sometime in the future,(It's based on SD1.5)

When I wanted to create a Persian girl in Stable Diffusion They would look more Indian than Persian so I decided to train my own LoRa to have better results.

It is not perfect and I haven't tested it a lot but in my limited testing I got relatively good results.

The dataset was trained of public pictures of random Persian girls and some of my friends with their permission.

Edit: If you see artifacts on the face, lower the weight and put it from 0.6 to 0.8, if this doesn't fix it send the output to image to image then use empty prompt and low Denoising strength to fix those issues, I hope to find a way to make it better in future versions.