Improved Diffusion is OpenAI’s official open-source release for the paper “Improved Denoising Diffusion Probabilistic Models,” hosted on GitHub with about 3,900 stars. It is the codebase that introduced several techniques (learned variances, cosine noise schedules and faster sampling) that became standard in modern image generation.
The repository contains the full training and sampling code for the improved DDPM models, including ImageNet-class diffusion models at multiple resolutions and the configurations used in the paper. Its contributions (predicting the variance alongside the mean, the cosine beta schedule, and importance-sampled training objectives) directly influenced later systems from OpenAI and the wider field, and this repo pairs with the companion guided-diffusion release. Pretrained checkpoints are provided so researchers can sample from the models or fine-tune them without repeating the original training runs. The codebase is a research release and was last updated in 2024.
Improved Diffusion is fully free and open source under the MIT license. There is no paid version, hosted service or usage restriction beyond the license terms.
Improved Diffusion is for generative modeling researchers, students and engineers who want to study, reproduce or build on the diffusion techniques that underpin today’s image generation systems.
The improvements in this release, by OpenAI researchers Alex Nichol and Prafulla Dhariwal, came from a systematic study of the original DDPM design: they showed that learning the reverse-process variance improves log-likelihood and sample quality, that a cosine noise schedule preserves signal longer than the original linear schedule, and that respacing the sampling steps trades a small amount of quality for large speedups. These ideas quickly became defaults across the field: the companion guided-diffusion codebase built classifier guidance on top of them, and the same family of models underpinned OpenAI’s GLIDE and DALL-E 2 image generation systems, which is why this compact research repository remains a reference point for anyone implementing diffusion models from scratch.