OpenSTL

Why we included this project

OpenSTL collects the main methods for spatio-temporal predictive learning in one codebase, covering tasks that run from synthetic moving objects to human motion, driving scenes, traffic flow, and weather forecasting. For researchers comparing video prediction approaches, the model zoo provides unified training and evaluation scripts, so baselines can be reproduced and cited without rebuilding each one from scratch. The framework is organized into core, algorithm, and interface layers, and offers both a PyTorch Lightning and a naive PyTorch implementation, letting teams work in whatever style suits them. An accompanying NeurIPS paper documents the benchmark setup, which is handy for write-ups that need to position a new method against established ones. If you are scoping a forecasting project and want to know which existing models actually work before building anything, this is a practical map of the field.