3/ V-JEPA – vision models trained on a feature prediction objective using 2 million videos; relies on self-supervised learning and doesn’t use pretrained image encoders, text, negative examples, reconstruction, or other supervision sources.

→ View original post on X — @dair_ai