This gave me some fits before I realized SDXL is a less-forgiving model to train than Flux or even SD 1.5. I had to fix some captions to get better results. v1.0 will give you a more consistent Priya, but it's also over-trained - this version is better at adhering to the prompt and the base model. Dim 32, alpha 16 (v1 is 32/24)