With the reinforcement learning of ruffles and frills on bottoms, the model can apply it in other areas. It is one of the reasons natural langue models such as XL and SD3 are harder to train the CLIP but can produce more specific request. Strangely enough we seem to like to refine those models with the old Danbooru and WD14 tags (1girl, ruffles, boy shorts vs a "a female on the beach wearing a frilly bikini and no top"