Denoising diffusion implicit models (DDIMs) are a more efficient class of iterative implicit probabilistic models that achieve high-quality image generation without adversarial training. DDIMs share the same training procedure as denoising diffusion probabilistic models (DDPMs), but they allow for faster sampling by using non-Markovian diffusion processes. These processes enable the generation of high-quality samples in significantly less time compared to DDPMs, with speedups of 10× to 50× in wall-clock time. DDIMs also enable semantically meaningful image interpolation directly in the latent space and can reconstruct observations with very low error. DDIMs are implicit probabilistic models that produce high-quality samples by leveraging non-Markovian diffusion processes, which allow for shorter generative Markov chains. This results in a significant increase in sample efficiency with only a minor cost in sample quality. The generative process of DDIMs is based on a variational inference objective that is equivalent to the DDPM objective for certain weights, allowing for the use of the same neural network with different diffusion processes. In experiments, DDIMs outperform DDPMs in terms of image generation when fewer iterations are considered, achieving higher sample quality with fewer steps. DDIMs also exhibit sample consistency, meaning that generated images with the same initial latent variable have similar high-level features regardless of the generation trajectory. This allows for direct interpolation in the latent space, which is not possible with DDPMs due to their stochastic generative process. DDIMs can also be used to encode samples that reconstruct them from the latent code, a capability not available with DDPMs. Additionally, DDIMs are related to neural ordinary differential equations (ODEs) and can be used to encode and decode data in a manner similar to ODEs. This makes DDIMs a versatile model for generating high-quality samples efficiently and performing meaningful interpolations in the latent space.Denoising diffusion implicit models (DDIMs) are a more efficient class of iterative implicit probabilistic models that achieve high-quality image generation without adversarial training. DDIMs share the same training procedure as denoising diffusion probabilistic models (DDPMs), but they allow for faster sampling by using non-Markovian diffusion processes. These processes enable the generation of high-quality samples in significantly less time compared to DDPMs, with speedups of 10× to 50× in wall-clock time. DDIMs also enable semantically meaningful image interpolation directly in the latent space and can reconstruct observations with very low error. DDIMs are implicit probabilistic models that produce high-quality samples by leveraging non-Markovian diffusion processes, which allow for shorter generative Markov chains. This results in a significant increase in sample efficiency with only a minor cost in sample quality. The generative process of DDIMs is based on a variational inference objective that is equivalent to the DDPM objective for certain weights, allowing for the use of the same neural network with different diffusion processes. In experiments, DDIMs outperform DDPMs in terms of image generation when fewer iterations are considered, achieving higher sample quality with fewer steps. DDIMs also exhibit sample consistency, meaning that generated images with the same initial latent variable have similar high-level features regardless of the generation trajectory. This allows for direct interpolation in the latent space, which is not possible with DDPMs due to their stochastic generative process. DDIMs can also be used to encode samples that reconstruct them from the latent code, a capability not available with DDPMs. Additionally, DDIMs are related to neural ordinary differential equations (ODEs) and can be used to encode and decode data in a manner similar to ODEs. This makes DDIMs a versatile model for generating high-quality samples efficiently and performing meaningful interpolations in the latent space.