Tried «new» (at least for me) way of training — setting high batch size (started with 16) and then lowering it to get more fine details (ended on 8 batch size). Result is for you to decide, but i think it's one of the best.

Trained on personally selected set, in addition I edited some images to remove unwanted artefacts like lips closed when they shouldn't be (result is arguable, but better than original)