Unsupervising Vision via Object-Centric World Models