Meta releases V-JEPA

Meta released V-JEPA, a video model built on the joint-embedding predictive architecture the company had introduced for still images the previous year with I-JEPA. Rather than generating video or reconstructing masked pixels directly, as diffusion and other generative video models do, V-JEPA was trained to predict the abstract representation of masked spatio-temporal regions from the surrounding context — an approach Yann LeCun, Meta’s chief AI scientist, had long argued was closer to how animals build world models than pixel-level prediction.

Meta reported that this non-generative training was substantially more compute-efficient than generative alternatives and that V-JEPA performed well at distinguishing fine-grained object interactions and short actions, over clips of up to roughly ten seconds, when its learned representations were evaluated on downstream tasks such as action recognition — without being fine-tuned on labelled video for those tasks. The model was trained entirely on unlabelled video, consistent with self-supervised approaches Meta had been pursuing across modalities.

Meta released the model under a non-commercial Creative Commons licence alongside research code, allowing academic use and further study but not commercial deployment. The release was framed explicitly as a step in LeCun’s longer-running research programme toward what he called more human-like machine intelligence — world models that predict at a level of abstraction rather than at the pixel level — a position that put Meta’s video research agenda in more direct contrast with the generative-video approach being pursued by OpenAI, Google DeepMind and video-generation startups over the same period.