Goku is a video generation foundation model that achieves industry-grade performance in joint image and video synthesis. It is built on rectified flow Transformers, a generative modeling approach that connects noise to data through ordinary differential equations. The project provides model weights, training code, and inference utilities for the video generation system.