“The VJPA 2.1 model, trained only on video prediction, learned representations that can predict depth from a single image better than the DINO-v3 model.”