Training Face Models

[alphanumeric-characters]

Just wondering if there's any tips or information regarding face model training? I know how to do it - rather I'm wondering what best practice is in terms of preparing a dataset for training. Am I better off using a select few high quality images that are visually coherent to maintain consistency? Or should I use a variety of images that include different expressions, lighting, angles etc in order to build a much more flexible / adaptable model? I imagine a single img2img swap would benefit from a dataset trained on images visually similar to the target image, however I'm coming at it from an img2vid workflow as well thus trained models must have flexibility.