V-JEPA

V-JEPA (Video Joint-Embedding Predictive Architecture) is a self-supervised method for learning visual representations from video, developed at Meta FAIR. The paper "Revisiting Feature Prediction for Learning Visual Representations from Video" (arXiv:2404.08471) was released on 15 February 2024. Authors: Adrien Bardes, Quentin Garrido, Jean Ponce, Xinlei Chen, Michael Rabbat, Yann LeCun, Mahmoud Assran, Nicolas Ballas.

V-JEPA is trained solely with a feature prediction objective: without pretrained image encoders, text, negative examples or pixel reconstruction. It extends the I-JEPA (2023) approach to the video domain and is a direct predecessor of V-JEPA 2.

Architecture and data

Vision Transformer backbone (ViT-L and ViT-H/16). Pre-trained on a collection of 2 million videos (VideoMix2M) gathered from public video datasets. Representations are evaluated on downstream image and video tasks without adapting the model's parameters (frozen backbone).

Results

The largest model, a ViT-H/16 trained only on videos, reaches: 81.9% top-1 on Kinetics-400, 72.2% on Something-Something v2 and 77.9% on ImageNet-1K — with a frozen backbone. This shows that learning by predicting video features yields versatile visual representations that perform well on both motion- and appearance-based tasks.

Position in the JEPA family

V-JEPA is the first JEPA-family model trained on video, the successor of I-JEPA (2023) and the direct predecessor of V-JEPA 2 (2025), which scales the approach to more than one million hours of video and adds an action-conditioned variant for robotic planning.