This paper presents improvements to Denoising Diffusion Probabilistic Models (DDPMs), showing that they can achieve competitive log-likelihoods while maintaining high sample quality. The authors demonstrate that learning variances of the reverse diffusion process allows sampling with fewer forward passes, which is important for practical deployment. They also compare DDPMs and GANs in terms of distribution coverage and find that diffusion models cover a larger portion of the target distribution. The study shows that sample quality and likelihood scale smoothly with model capacity and training compute, making DDPMs easily scalable. The authors release their code at https://github.com/openai/improved-diffusion. The paper introduces a hybrid objective that combines the variational lower-bound (VLB) with a simplified objective from Ho et al. (2020), which leads to better log-likelihoods and reduces gradient noise. They also propose a cosine noise schedule that improves sample quality compared to the linear schedule used in previous work. The authors find that using a cosine schedule allows for faster sampling with fewer diffusion steps while maintaining high sample quality. They also show that the hybrid objective outperforms the VLB objective in terms of log-likelihood, while maintaining similar FID scores. The paper also explores the scalability of DDPMs with increasing model size and training compute, finding that performance improves predictably as training compute increases. The authors compare their results to other likelihood-based models and find that DDPMs are competitive with the best conventional methods in terms of log-likelihood. They also show that DDPMs can match the sample quality of GANs while achieving better mode coverage as measured by recall. The paper concludes that DDPMs are a promising direction for future research, as they combine good log-likelihoods, high-quality samples, and reasonably fast sampling with a well-grounded, stationary training objective that scales easily with training compute.This paper presents improvements to Denoising Diffusion Probabilistic Models (DDPMs), showing that they can achieve competitive log-likelihoods while maintaining high sample quality. The authors demonstrate that learning variances of the reverse diffusion process allows sampling with fewer forward passes, which is important for practical deployment. They also compare DDPMs and GANs in terms of distribution coverage and find that diffusion models cover a larger portion of the target distribution. The study shows that sample quality and likelihood scale smoothly with model capacity and training compute, making DDPMs easily scalable. The authors release their code at https://github.com/openai/improved-diffusion. The paper introduces a hybrid objective that combines the variational lower-bound (VLB) with a simplified objective from Ho et al. (2020), which leads to better log-likelihoods and reduces gradient noise. They also propose a cosine noise schedule that improves sample quality compared to the linear schedule used in previous work. The authors find that using a cosine schedule allows for faster sampling with fewer diffusion steps while maintaining high sample quality. They also show that the hybrid objective outperforms the VLB objective in terms of log-likelihood, while maintaining similar FID scores. The paper also explores the scalability of DDPMs with increasing model size and training compute, finding that performance improves predictably as training compute increases. The authors compare their results to other likelihood-based models and find that DDPMs are competitive with the best conventional methods in terms of log-likelihood. They also show that DDPMs can match the sample quality of GANs while achieving better mode coverage as measured by recall. The paper concludes that DDPMs are a promising direction for future research, as they combine good log-likelihoods, high-quality samples, and reasonably fast sampling with a well-grounded, stationary training objective that scales easily with training compute.