That's interesting because I've always found the AI portraits to look extremely unrealistic. However, they were probably using a very specifically trained model for realistic portraits where most of my experiments with it have been using more general models. I could almost never get it to generate a convincing portrait.