Download
VJEPA2 is a next-generation self-supervised learning framework for video that extends the “predict in representation space” idea from i-JEPA to the temporal domain. Instead of reconstructing pixels, it predicts the missing high-level embeddings of masked space-time regions using a context encoder and a slowly updated target encoder. This objective encourages the model to learn semantics, motion, and long-range structure without the shortcuts that pixel-level losses can invite. The architecture is designed to scale: spatiotemporal ViT backbones, flexible masking schedules, and efficient sampling let it train on long clips while remaining stable. Trained representations transfer well to downstream tasks such as action recognition, temporal localization, and video retrieval, often with simple linear probes or light fine-tuning. The repository typically includes end-to-end recipes—data pipelines, augmentation policies, training scripts, and evaluation harnesses.
Features
- Predictive learning in embedding space for masked space-time regions
- Context and EMA target encoders for stable self-supervised training
- Spatiotemporal ViT backbones with scalable masking strategies
- Strong transfer with linear probes on standard video benchmarks
- Efficient training without pixel reconstruction or negative pairs
- Turnkey data pipelines and evaluation scripts for rapid reproduction
Project Samples
