Learning Priors of Human Motion With Vision Transformers