V-JEPA by Meta

V-JEPA (Video Joint Embedding Predictive Architecture) is a self-supervised learning method for video developed by Meta AI Research (FAIR). The official PyTorch codebase provides pre-trained Vision Transformer (ViT-L and ViT-H) models trained on the VideoMix2M dataset using an unsupervised feature prediction objective. V-JEPA produces versatile visual representations that perform well on downstream video and image tasks without adapting the model's parameters, requiring only a lightweight task-specific attentive probe. The method does not use pretrained image encoders, text, negative examples, human annotations, or pixel-level reconstruction; instead, a predictor makes predictions in latent space, and a conditional diffusion model can decode these predictions to interpretable pixels. Evaluations on Kinetics-400, Something-Something v2, ImageNet-1K, Places205, and iNaturalist 2021 show competitive accuracy (e.g., 82.0% on K400 with ViT-H).