VideoMAE

Official PyTorch implementation of VideoMAE, a masked autoencoder framework for self-supervised video pre-training. The repository provides training scripts, pretrained models, and utilities to pretrain and finetune video representations on standard datasets (e.g., Kinetics-400). VideoMAE targets data-efficient learning for video understanding via masked autoencoding, developed by Nanjing University and Tencent AI Lab.

AI named VideoMAE in August 2026.