Trained with a set of cropped images with this type of breasts. No faces in the training data, mostly just torso and breasts. I used triggers in the data but they don't make much difference so most preview images is done without a special trigger word.