SR3 (Super-Resolution via Repeated Refinement) is a novel approach to image super-resolution that leverages denoising diffusion probabilistic models. It performs super-resolution through a stochastic iterative denoising process, starting with pure Gaussian noise and refining it using a U-Net model trained on denoising at various noise levels. SR3 demonstrates strong performance on super-resolution tasks at different magnification factors, both on faces and natural images. Human evaluation on an 8× face super-resolution task on CelebA-HQ shows that SR3 achieves a fool rate close to 50%, outperforming SOTA GAN methods which achieve a fool rate of at most 34%. SR3 is also effective in cascaded image generation, where generative models are chained with super-resolution models, yielding competitive FID scores on ImageNet. The key contributions of SR3 include adapting denoising diffusion models to conditional image generation, achieving high-quality super-resolution across various magnification factors, and demonstrating the effectiveness of cascaded models for efficient inference and unconditional generation.SR3 (Super-Resolution via Repeated Refinement) is a novel approach to image super-resolution that leverages denoising diffusion probabilistic models. It performs super-resolution through a stochastic iterative denoising process, starting with pure Gaussian noise and refining it using a U-Net model trained on denoising at various noise levels. SR3 demonstrates strong performance on super-resolution tasks at different magnification factors, both on faces and natural images. Human evaluation on an 8× face super-resolution task on CelebA-HQ shows that SR3 achieves a fool rate close to 50%, outperforming SOTA GAN methods which achieve a fool rate of at most 34%. SR3 is also effective in cascaded image generation, where generative models are chained with super-resolution models, yielding competitive FID scores on ImageNet. The key contributions of SR3 include adapting denoising diffusion models to conditional image generation, achieving high-quality super-resolution across various magnification factors, and demonstrating the effectiveness of cascaded models for efficient inference and unconditional generation.