V-JEPA 2 is a self-supervised learning framework that trains video encoders using masked latent feature prediction on internet-scale video data.
The project addresses the challenge of learning robust video representations without labeled data. It uses a self-supervised approach where an encoder and predictor network are trained to predict masked latent features from video, enabling the model to learn physical world understanding and temporal dynamics. V-JEPA 2 achieves state-of-the-art performance on motion understanding and human action anticipation tasks. The framework also includes V-JEPA 2-AC, a variant that can be post-trained as a latent action-conditioned world model using robot trajectory data to solve manipulation tasks without environment-specific data collection or task-specific training.
V-JEPA 2.1 extends the approach with improvements focused on learning temporally consistent dense features through a dense predictive loss where all tokens contribute to training, deep self-supervision applied at multiple encoder layers, multi-modal tokenizers for images and videos, and benefits from model and data scaling. Developers should consider this tool if they need pre-trained video encoders for downstream tasks involving motion understanding, action prediction, or robotics applications. The framework is particularly suited for projects that can leverage internet-scale video data during pre-training and want to avoid task-specific labeled data collection.
The codebase is actively maintained with multiple model variants released. The project includes official PyTorch implementations with accompanying research papers and blog posts documenting the approach and results. The team has published detailed documentation of the training methodology, architecture choices, and performance benchmarks across different downstream tasks.
