video-diffusion-pytorch

Why we included this project

Jonathan Ho's video diffusion paper was one of the first to extend denoising diffusion models from still images to short clips, and this PyTorch port captures the core design in a compact package. The space-time factored U-Net and the Gaussian diffusion trainer come as straightforward classes, so you can train a small video model with a few lines of code instead of reimplementing the architecture yourself. It's best treated as a learning tool and a starting scaffold: the author reports it converges on simple synthetic tasks like moving MNIST, and points anyone chasing real text-to-video results toward the more active imagen-pytorch repository. For someone who wants to read a working video diffusion pipeline end to end, this is a good place to start.