A high quality training set is super important, clean sharp images with good lighting are ideal. I mention this to the customers in the app, but this is out of my control. If you are choosing your own data, then this may not be as much of a problem.
Same goes with images being representative of them. Eg: not too much make up, not using old photos, not using sunglasses etc.
As for training, regularization or class images help a lot. This shows the dreambooth model other images of people that are not your subject and helps it differentiate.
Number of training epochs is something to experiment with. That number will change with the underlying base model you use, with the number of training images used, with learning rate etc.
Both SD1.5 and SDXL models are good, have their pros and cons, but that's a longer discussion.