vjepa2

Why we included this project

Most video models describe what is in a frame, but V-JEPA 2 is built for understanding what happens across time. It is Meta FAIR's self-supervised video encoder, trained on internet-scale video without labels, and it reports state-of-the-art results on motion understanding and human action anticipation. The same codebase includes V-JEPA 2-AC, a world model post-trained on a small amount of robot trajectory data that handles manipulation tasks without task-specific training or calibration. Since training is self-supervised, you can adapt the encoder to your own video data without a large labeling effort, and the released PyTorch code and checkpoints give you something runnable rather than a paper to implement yourself. A 2.1 release updates the training recipe to produce denser, more temporally consistent features.