Denoising Diffusion Implicit Models (DDIMs) are introduced as an efficient alternative to Denoising Diffusion Probabilistic Models (DDPMs) for generating high-quality images. DDIMs are a class of iterative implicit probabilistic models that share the same training procedure as DDPMs but require fewer steps to produce samples, making them significantly faster. The key innovation is the use of non-Markovian diffusion processes, which allow for shorter generative Markov chains while maintaining the same training objective as DDPMs. This results in samples that are 10× to 50× faster to generate compared to DDPMs, with only a minor loss in sample quality. DDIMs also exhibit "consistency" in the latent space, allowing for meaningful image interpolation and reconstruction with low error. The paper demonstrates these benefits through empirical evaluations on various datasets, showing that DDIMs can achieve high-quality samples with significantly reduced computational time.Denoising Diffusion Implicit Models (DDIMs) are introduced as an efficient alternative to Denoising Diffusion Probabilistic Models (DDPMs) for generating high-quality images. DDIMs are a class of iterative implicit probabilistic models that share the same training procedure as DDPMs but require fewer steps to produce samples, making them significantly faster. The key innovation is the use of non-Markovian diffusion processes, which allow for shorter generative Markov chains while maintaining the same training objective as DDPMs. This results in samples that are 10× to 50× faster to generate compared to DDPMs, with only a minor loss in sample quality. DDIMs also exhibit "consistency" in the latent space, allowing for meaningful image interpolation and reconstruction with low error. The paper demonstrates these benefits through empirical evaluations on various datasets, showing that DDIMs can achieve high-quality samples with significantly reduced computational time.