This model was trained on 35 images of various types of luchador masks, worn by both males and females. All of the sample images just use "luchador_mask" in the prompt, and with a weight of 1. Even though there were only a few images with ear or horn-like elements in the training set, there does seem to be a tendency to add those details. (I'm guessing superhero masks in the various checkpoints are influencing this.) You can reduce this issue by adding things like "horns" and "cat ears" in the negative prompt.