Denoising Diffusion Implicit Models (DDIM) are a class of generative models introduced in 2021 by Jiaming Song, Chenlin Meng, and Stefano Ermon from Stanford University. They extend the diffusion model framework by redefining the sampling process as an implicit probabilistic model, allowing for faster generation of high-quality samples from a trained diffusion model without requiring additional training or modification to the underlying network. DDIMs address a key limitation of standard diffusion models: the slow sampling process that typically requires thousands of iterative denoising steps to produce a single image. By formulating the reverse process as a non-Markovian chain, DDIMs enable the same trained model to sample with substantially fewer steps, often 10 to 50 times fewer, while retaining competitive output quality.
The core innovation of DDIM lies in a reformulation of the diffusion process's forward and reverse transitions. Standard denoising diffusion probabilistic models (DDPMs) assume a Markov chain where each step depends only on the previous state, necessitating a large number of steps to reconstruct a clean sample from Gaussian noise. DDIM instead observes that the training objective of a diffusion model is consistent with a broader family of non-Markovian forward processes, which can be designed to allow for deterministic or semideterministic reverse processes. By parameterizing these processes with a free variable, DDIM constructs a class of generative models that yields the same training loss but more flexible sampling trajectories. This permits a deterministic mapping from noise to data, enabling the generation of samples in a straightforward and reproducible way, a property absent from the original stochastic formulations.
The practical benefit of this non-Markovian perspective is the ability to sample using a reduced timestep schedule. While a DDPM typically uses 1000 to 2000 denoising steps, DDIM can produce valid images using just 20, 50, or 100 steps in many cases. The model's architectural components remain identical to those in DDPM - a neural network, usually a U-Net with attention layers, predicts the noise component at each step. The difference is probabilistic process, which allows skipping intermediate steps in the denoiser schedule without introducing major reconstruction errors. As a result, DDIM offers a trade-off between sampling speed and sample fidelity; higher step counts yield high accuracy, while lower steps sacrifice some quality but remain visually coherent. This flexibility is exploitable for tasks such as inference-time guidance, where the same model can produce diverse samples under different schedules.
Beyond speed, DDIM enables a deterministic and invertible mapping between the latent noise and the generated data. This property supports interpolation in the noise space, where linearly interpolating between two samples in latent space yields semantically meaningful morphing in the resulting image space. This deterministic nature also allows for sampling like in generative adversarially trained models, for tasks such as image editing, inpainting, or reconstruction from partial observations, without the associated instability of training GANs. The implicit model view also places DDIM within the broader family of implicit probabilistic models, which define distributions via a simple deterministic generating function, linking to concepts from variational inference and normalizing flows.
The algorithmic development of DDIM contributed to the broader evolution of efficient diffusion-based Generative AI. Its sampling formula, often written as a linear combination of predicted noise and current sample with a scaled coefficient, became a building block for later works that addressed latent diffusion, text-to-image synthesis, and high-resolution generation. DDIM is a widely used denoising sampler in several reference implementations, including those associated with the Stable Diffusion model from ai-archived? - its step skip strategy and deterministic formulation remain a standard of comparison for subsequent work in the field, such as the denoising diffusion GANs and consistency models.
Relationship to Diffusion Probabilistic Models
DDIM builds upon the theoretical foundation provided by diffusion probabilistic models, notably the work of Sohl-Dickstein et al. (2015) and Diederik Ho et al. (2020). The standard DDPM randomly perturbs data with no overloaded with Gaussian noise over 1000 steps and learns to reverse this process by estimating the noise. While effective, sampling from DDPM draws a new random noise sample at each reverse step. DDIM instead defines a family of forward processes such that the reverse process can be approximately deterministic. A key theoretical insight is that the loss function involves only the KL divergence between the reverse and forward processes chosen to depend on the score estimator; alternate forward processes are possible without affecting training, allowing the flexibility in sampling. DDIM then constructs the 'implicit' chain to match these statistical marginals, resulting in a new sampling rule.
A direct consequence of this flexibility is that DDIM is a separate family of models, but the same training procedure works for both. A user must keep the same network, but at sampling time, they can choose to use the DDPM schedule or the DDIM schedule, depending on time constraints. The parameters of the DDIM formula are determined by the alpha (α) and sigma (σ) schedule of the noise, which are often precomputed as the diffusion coefficients. The core equation of DDIM relates the predicted denoised sample (x0) and the current latent (xt) to produce the next latent x(t-1) using the formula: x(t-1) = sqrt(α(t-1)) x0 + sqrt(1 - α(t-1) - σ(t)^2) ε_θ(xt) + σ(t) z, where z is optional noise and ε is the noise prediction. When σ is set to zero, the process is fully deterministic, yielding a family called the "inhomogeneous" DDIM, which corresponds to the deterministic annealed Langevin dynamics.
The work originates at BAIR (Berkeley AI Research) (in the form of Song and Ermon affiliations, although they were at Stanford at the time) and is related to the earlier score-based generative models and noiseless networks. The DDIM formulation is also closely tied to Neural network approximators of score functions, as the noise predictors are Gaussians with the goal of estimating the score term ensures that the reverse of diffusion steps captures data manifold.
Advantages and Limitations
Practical benefits of DDIM include a sample-time controllability, reproducibility (because the random seed accounts for the input latent), and extrapolation. It provides an effective, joint estimate of the image based on a fixed number of denoising steps, making it possible to utilize the same model on resource-constrained hardware. However, DDIM is based on the assumption of a Gaussian error at each step; in practice, the finite step approximation limits the quality compared to the full SDE sampler. As a result, the performance drops noticeable at moderate to low step counts, with severe artifacts for very few steps, leading to deformed anatomy or blurs. Nevertheless, in the era of Large language model-adjacent text-to-image models, the deterministic decrement characteristic is useful for building pipelines of sampling during Deep learning inference.
The explicit and denoising mechanism also helps in the construction of a common architecture of DDPM derivatives, such as the Denoising Diffusion Implicit Models used in Stable Diffusion and High-Resolution image synthesis. The stable diffusion approach uses a reduced latent space and the DDIM sampler as an efficient solver for the reverse diffusion in that latent core. This formed the basis of the widely-used diffusers (library) package, and the actual code for DDIM is often bundled with all other diffusion models, making its accessibility high.
Sampling speed and quality trade-off
DDIM accepts a small and manageable two-step schedule that was on the standard of the original diffusion. The number of steps is a hyperparameter that can be adjusted at inference to balance the quality of samples. With 20 steps, a DDIM is capable of producing images similar in quality to a 1000-step DDPM in many contexts, but produces more pronounced visual artifacts such as blurred regions if you reduce it to 10 steps. The use of DDIM in a certain module, such as the sharp sampler from Latent Diffusion models, remains using 50 steps to competitive metrics. The schedule often of how to traverse the initial noise to the output is - for example, a straight-line interpolation between the starting latent and the final mean noise? - that the output tends to be of moderate quality at early and late steps.
Compared to methods that use the adaptive sampling scheduler, such as the AlignSAM ArrayList or the Conclément implicits, DDIM does not retrain the network but rather trains the denoiser, the same as a classifier, and the proposed is able to obtain good outcomes.
Impact and related work
Following publication at the International Conference on Learning Representations (ICLR) 2021, DDIM made an impact on the wider generative modeling field. It introduced the concept of deterministic sampling for diffusion processes, which is crucial for understanding the space of latent variables and enables the use of diffusion models in downstream tasks like Latent Diffusion Models (Rombach et al., 2022), which lock the based on the ID of the of embeddings and the DDIM sampler. Many subsequent articles cite the approach in “Fisher merger” and “authors.” The training objective is the same as DDPM but its sample handling leads to a better latent of the GAN style. The reverse approach in DDIM has also been used in the image-to-image translation and the family of posterior, which is known in the inverse set up.
The existence of these findings paved the way for powerful sampling algorithms, such as the DPM-solver (2022) that is a deterministic method using integration analysis, and the denoising diffusion GAN model (2022) to model the step process with adversarial steps. These modern approaches often combine the DDIM base with continuous test that converge, sometimes obtaining a sample in one iteration. The philosophy of the non-Markov viewpoints is also how subsequent models improved. DDIM's deterministic dynamics are an essential tool in today's AI system labs, often in the LangChain for instance.
In addition, the theory of implicit probability models research connected with the flow and the rectified flow, which are used in computer engines as a design tool. The paper has been cited highly, ranking as one of the foundational papers for the part of the open-source AI field of generating video effects.
Overall, DDIM's importance lies in the high step, its elegance, and the bridging between the Computational and the Practical. It makes the high-quality generation accessible to a diversity of users, and it remains a standard sample in any modern text-to-image or diffusion-based pipeline, including ones from Google DeepMind-adjacent, OpenAI-adjacent stable dict and Adobe. For all practical (internal) in 2025, DDIM is used as the classic sampling engine.
Sources
- Song et al., 2021, "Denoising Diffusion Implicit Models" (online preprint)
- Ho et al., 2020, "Denoising Diffusion Probabilistic Models"