DDIM

The Problem

The original Denoising Diffusion Probabilistic Model (DDPM) generated incredible images, but it was practically unusable for real-world applications due to its speed. Because the mathematical formulation of DDPM relied on a strict Markov chain, the reverse (generation) process had to exactly mirror the 1,000 tiny steps of the forward (noising) process. To generate one image, you had to run a heavy U-Net neural network 1,000 times sequentially.

The Idea

The authors realized that the forward noising process didn't have to be a Markov chain (where step $t$ strictly depends only on step $t-1$). They rewrote the mathematical framework as "non-Markovian," where the noise added at any step could depend directly on the original image $x_0$. Crucially, they proved that you could use the exact same trained DDPM model, but sample from it using this new formulation. This new sampler allowed them to safely "skip" steps during generation, turning a 1,000-step process into a 50-step process with almost no loss in quality.

How It Works

DDIM creates a generalized class of non-Markovian inference processes.

In DDPM, the sampling equation includes a random noise term at every step (hence "Probabilistic"). The DDIM formulation introduces a parameter $\sigma$ that controls the variance of this noise.

If you set $\sigma = 0$, the process becomes entirely deterministic (hence "Implicit"). Because there is no randomness injected at each step, the trajectory from pure noise to the final image becomes a fixed curve. This smooth, predictable trajectory means you can take much larger "jumps" across the curve (skipping steps in time) without drifting off course.

Why It Mattered

DDIM made diffusion models commercially viable. By cutting generation time by 10x to 50x, it allowed these models to be deployed in production APIs. Furthermore, because the DDIM sampling process is deterministic, it meant that the exact same initial noise vector would always produce the exact same image. This unlocked capabilities like image editing and interpolation in latent space, which were impossible with the noisy DDPM sampler.

What Came After

DDIM became the standard solver/sampler for almost all diffusion models (including Stable Diffusion). It sparked a whole sub-field of research into "ODE solvers" for diffusion models (like DPM-Solver and Euler samplers) which framed the generation process as solving a differential equation, eventually bringing sampling steps down to the single digits.