Trained with 57 images, 10 epoch

Slightly lower offset

Regarding the logo problems, most of my dataset are cleaned up already, lora is very sensitive to text. However I think most SDXL model are heavily influenced by this that if you try to generate NIKKE cahracter the logo will pop up occasionally. Maybe a bit of negative prompt can help.

If you like my work please consider supporting me: ko-fi.com/suzushi2024

I am also building a civit ai alternative for you to share your model and images, it will purely be model hosting at the moment with less restrictions: https://miyukiai.com/