Learning Vision from Models Rivals Learning Vision from Data