PyTorchVideo is Meta AI’s deep learning library for video understanding research, providing modular components, pretrained models and benchmarks for tasks such as action recognition, video classification and temporal detection. Built on PyTorch, it aims to make video research as reproducible as image research became with torchvision-style tooling. The project has around 3,600 stars on GitHub and is Apache-2.0 licensed.
The library covers the full video pipeline: data loading and decoding for datasets such as Kinetics, Something-Something and AVA; a model zoo implementing architectures including SlowFast, X3D, MViT, C2D and I3D with pretrained weights; and training and evaluation recipes with standard metrics. Components are deliberately composable โ data transforms, models, losses and accelerators can be mixed and extended โ and the same codebase supports research experiments and deployment-oriented workflows. PyTorchVideo also underpins Meta’s own video research and integrates with the wider PyTorch ecosystem, including TorchHub for one-line model loading.
PyTorchVideo is completely free and open source under the Apache-2.0 license. The only cost is GPU compute for training and inference.
Video-understanding researchers and engineers building products around action recognition, sports analytics, content moderation or surveillance who want proven architectures and benchmarks rather than starting from raw PyTorch.
The project also provides a dedicated data API for building custom video datasets, accelerator support for efficient training on modern hardware, and reference implementations that have been used across Meta’s published video research, making it a dependable starting point for new video-understanding projects.