V-JEPA 2 (2025) › Technical Innovations

Several technical advances enabled V-JEPA 2's world-modeling capabilities. The visual encoders employed transformer-based architectures adapted for visual inputs. This enabled the model to process spatial relationships and extract multi-scale features from images. These encoders learned representations that captured both local visual details and global scene structure. This provided rich inputs for the prediction mechanisms.

The prediction architecture learned to forecast future visual representations in the embedding space, rather than predicting raw pixels. This target allowed the model to focus on semantic changes in scenes rather than pixel-level details. By learning to predict how representations would evolve, the model captured the essential dynamics of physical interactions: how objects move, how forces affect motion, and how actions produce consequences.

Joint embedding spaces unified the representation of current observations, predicted future states, and actual future states. This enabled the model to learn from prediction errors, adjusting its understanding of physical dynamics when predictions failed. The embedding space captured abstract relationships about how scenes evolve, encoding knowledge about physics and causality in a form that could be learned from visual data.

The training process used contrastive learning techniques, where the model learned to make predicted future representations similar to actual future representations when predictions were correct, and dissimilar when predictions were wrong. This self-supervised approach enabled learning from vast amounts of unlabeled video data, discovering patterns about physical dynamics without requiring manual annotations or text descriptions.