DALLE2-pytorch

Why we included this project

DALL-E 2 generates images in two stages: a diffusion prior turns a CLIP text embedding into an image embedding, and a decoder renders that embedding into pixels. This repo implements the whole pipeline in PyTorch, and the author chose to build out the diffusion prior variant specifically, so what you get is working code rather than just the paper's equations. It's genuinely trainable end to end, and outside groups have used it to train real diffusion priors for CLIP and StyleGAN pipelines; one team scaled training across 800 GPUs using the bundled scripts. The README is candid that newer text-to-image models have since overtaken it, but if you want to study the two-stage design or retrain a variant yourself, this is a solid starting point.