imagen-pytorch

Why we included this project

This repo is a faithful PyTorch port of Google's Imagen, the cascading diffusion text-to-image model that outperformed DALL-E 2 when it was published. It strings together a pretrained T5 encoder, a memory-efficient U-Net, and dynamic clipping for classifier-free guidance, and the code is modular enough to follow how those pieces connect without rereading the paper line by line. That makes it a solid study and experimentation base for researchers and engineers who want to train or adapt a diffusion model of their own. It is not a polished, plug-and-play service, so expect to bring your own training setup, but as a reference implementation to build on it is hard to beat.