jepa

Why we included this project

Meta FAIR's official V-JEPA codebase trains visual representations by watching unlabeled video. Instead of reconstructing pixels, the model predicts missing spatiotemporal regions in latent space, which keeps pretraining free of text, annotations, negative examples, and pixel-level reconstruction. The resulting frozen ViT backbones transfer to both video and image tasks, where a lightweight attentive probe reaches strong accuracy on Kinetics-400, Something-Something-v2, and ImageNet. If you build video understanding systems with limited labeled data, the bundled ViT-Large and ViT-Huge checkpoints plus evaluation scripts give you a solid starting point. It's a research codebase built for reproducibility, so expect config-driven PyTorch training on CSV-based datasets rather than a ready-made inference API.