Denoising Diffusion Implicit Models

To understand DDIM, we first need to see why DDPM is slow. In DDPM, the forward process gradually adds Gaussian noise to a clean image $x_0$ over $T = 1{,}000$ steps, producing a sequence $x_1, x_2, \ldots, x_T$ that ends in pure noise. This is a Markov chain: each step $x_t$ depends only on the previous step $x_{t-1}$.

The generative process reverses this chain: starting from pure noise $x_T \sim \mathcal{N}(0, I)$, it denoises step by step until reaching a clean image $x_0$. Because the forward process has $T$ steps, the reverse process must also take $T$ steps — one neural network evaluation per step. There is no shortcut in a Markov chain.

A key property makes this possible: we can jump from $x_0$ to any step $x_t$ directly:

This is the crucial observation: the marginal distribution $q(x_t | x_0)$ at each step does not depend on the path taken to get there. Whether we arrived through 1,000 Markov steps or jumped directly, the statistics are the same. The DDPM training objective (predicting the noise $\epsilon$ from $x_t$) only depends on these marginals — not on the joint distribution $q(x_1, \ldots, x_T | x_0)$. This is the door that DDIM opens.

The Markov chain in DDPM is just one particular joint distribution $q(x_1, \ldots, x_T | x_0)$ that produces the marginals $q(x_t | x_0) = \mathcal{N}(\sqrt{\alpha_t}\, x_0, (1 - \alpha_t) I)$. But there are infinitely many other joint distributions that produce the exact same marginals.

Think of it like this: imagine you know the weather statistics for every month of the year. The Markov model says "today's weather depends only on yesterday." But you could also have a model where today's weather depends on yesterday and on the overall season — a non-Markovian model. Both can produce the same monthly averages, but the day-to-day transitions look very different.

DDIM exploits this freedom. It defines a family of non-Markovian forward processes, indexed by a vector $\sigma = (\sigma_1, \ldots, \sigma_T)$, where each $\sigma_t$ controls how much randomness enters at step $t$. The family is designed so that the marginals $q_\sigma(x_t | x_0)$ are identical to DDPM's for every $\sigma$. Since the training objective depends only on these marginals, the same pretrained model works for the entire family — no retraining needed.

The heart of DDIM is a single update equation that generates $x_{t-1}$ from $x_t$. Before we see the math, let's understand what it's doing.

The model $\epsilon_\theta^{(t)}(x_t)$ has been trained to predict the noise that was added to the clean image. Given this prediction, we can estimate what the clean image $x_0$ looks like. Then we "re-noise" this estimate to the correct level for step $t{-}1$. The equation blends three components: the predicted clean image, a correction term that points in the "noise direction," and an optional random noise term controlled by $\sigma_t$.

Think of it as a tug-of-war between "signal" and "noise." The first term pulls toward the clean image — like a compass pointing home. The second term adjusts the noise level so we land at the right position on the denoising schedule. The third term adds controlled randomness that is useful for diversity but slows convergence.

The parameter $\eta$ provides a convenient way to interpolate. Setting $\eta = 0$ gives DDIM (deterministic), $\eta = 1$ gives DDPM (stochastic), and values in between give a smooth continuum. This is achieved by defining $\sigma_t(\eta) = \eta \sqrt{(1-\alpha_{t-1})/(1-\alpha_t)} \sqrt{1 - \alpha_t / \alpha_{t-1}}$.

The second breakthrough is that we don't need to reverse all $T = 1{,}000$ steps. Since the training objective only depends on marginals $q(x_t | x_0)$, we can choose a subsequence $\tau = [\tau_1, \tau_2, \ldots, \tau_S]$ of the original $[1, \ldots, T]$ and run the generative process only on those steps. Instead of going through $x_{1000} \to x_{999} \to \cdots \to x_0$, we might jump $x_{1000} \to x_{900} \to x_{800} \to \cdots \to x_0$ — taking only 10 steps.

The update equation works exactly the same, just with $\alpha_{\tau_i}$ and $\alpha_{\tau_{i-1}}$ replacing $\alpha_t$ and $\alpha_{t-1}$. The model was trained on all 1,000 noise levels, so it can denoise at any of them. We're simply choosing to visit fewer waypoints on the journey from noise to image.

The experimental results are striking. On CIFAR-10, DDIM with 50 steps achieves an FID of 4.67, comparable to DDPM's 4.04 with 1,000 steps — a 20× speedup with minimal quality loss. With 20 steps, DDIM still scores 6.84, while DDPM collapses to 18.36 under the same budget.

Because DDIM is deterministic, the same initial noise $x_T$ always produces the same image $x_0$ — regardless of how many steps we use. This is called consistency: a 10-step and a 1,000-step generation from identical $x_T$ share the same high-level features (pose, color, composition), differing only in fine details.

This transforms $x_T$ into a proper latent code, similar to the latent space in GANs or variational autoencoders. We gain two powerful capabilities. First, interpolation: given two noise vectors $x_T^{(0)}$ and $x_T^{(1)}$, we can smoothly walk between them using spherical linear interpolation and generate all intermediate images. The transitions are semantically meaningful — faces morph smoothly, scenes blend naturally.

Second, encoding: by running the DDIM update in reverse (from $x_0$ to $x_T$), we can encode any image into the latent space, then decode it back with near-perfect reconstruction. The reconstruction error drops below $10^{-4}$ with 1,000 encoding steps. This is impossible with DDPM because its stochastic process destroys the one-to-one mapping.

There is a beautiful mathematical connection hidden beneath the surface. If we reparametrize by dividing each $x_t$ by $\sqrt{\alpha_t}$ and defining $\bar{\sigma}(t) = \sqrt{(1-\alpha_t)/\alpha_t}$, the DDIM update becomes a single Euler step of an ordinary differential equation:

This ODE perspective is powerful for two reasons. First, it means we can apply decades of research on numerical ODE solvers (Adams-Bashforth, Runge-Kutta) to improve sampling quality in fewer steps. Second, it gives us a reversible encoding: unlike DDPM, where stochasticity makes the mapping from $x_0$ to $x_T$ one-to-many, the DDIM ODE defines a bijection — every image has exactly one latent code, and vice versa. This makes DDIM a normalizing flow by another name, with the noise prediction network playing the role of the velocity field.

DDIM solved the practical barrier that stood between diffusion models and real-world deployment. Before DDIM, diffusion models were an academic curiosity — beautiful math, excellent sample quality, but too slow to use. After DDIM, diffusion models became commercially viable.

The impact was immediate and far-reaching. Every modern diffusion-based system — Stable Diffusion, DALL·E 2, Imagen, Midjourney — uses DDIM-style accelerated sampling as a fundamental building block. The deterministic latent space enabled image editing techniques like DDIM inversion, where real photos are encoded into latent space, edited, and decoded back. The ODE connection inspired DPM-Solver, which pushed the boundary further to 10–20 steps with higher-order ODE solvers. And the core idea — that a single trained model supports many inference strategies — directly inspired consistency models, which distill the entire trajectory into a one-step generator.