mocogan

Why we included this project

MoCoGAN is the reference implementation of a CVPR 2018 paper that splits video generation into two separate latent codes: content, which decides what appears in the frame, and motion, which decides how it moves. That split is what makes the model interesting, because you can hold one code fixed and vary the other, producing the same person doing different actions or different people doing the same action. The repo includes training code, a Docker image for setup, and examples trained on facial expressions, human actions, and a TaiChi dataset, so reproducing the paper's results is fairly straightforward. It is a research codebase, not a production service, and its real value is as a baseline for studying motion-content disentanglement or a starting point for extending video GANs.