I made a change in my training workflow so I re-did this Lora. I think it's a little better. What I've noticed with XL vs 1.5 and even Flux is the outcomes are influenced a lot by the checkpoint biases (and other Loras, presumably) and it's not as good at "filling in the blanks."So you may need more detail in the prompt; e.g. saying she has brown or dark skin, saying she has a certain hair length, etc.