It's a good concept but because it's trained with a girl in the image it has a big bias for sticking a girl into the shot. That means if you're using controlNet to add a person the model constantly wants to add a second person, usually in the center of the image. It would be better to train a version without people.