Scaling Language-Free Visual Representation Learning