Meta AI Releases V-JEPA 2: Open-Source Self-Supervised World Models for Understanding, Prediction, and Planning

Meta AI has unveiled V-JEPA 2, an open-source, scalable world model capable of learning from over 1 million hours of internet video and images to enhance visual understanding, future state prediction, and zero-shot planning. Building on the joint-embedding predictive architecture (JEPA), V-JEPA 2 employs self-supervised learning through a visual mask denoising objective, enabling the model to reconstruct masked spatiotemporal patches in a latent space, thereby focusing on scene dynamics while ignoring noise. To achieve this scale, Meta researchers developed techniques such as constructing a large dataset (VideoMix22