PAPERS-04-007 · Paper — Denoising Diffusion (DDPM)

The paper

The paper “Denoising Diffusion Probabilistic Models” is a landmark work that introduced a new way to generate images by learning to reverse a gradual noising process, showing that diffusion can rival traditional generative methods.

Before this paper

Before this work, GANs were the dominant image generators. They could produce impressive pictures but were notoriously unstable during training and often suffered from mode collapse, where only a few types of images were generated. Researchers were looking for a more reliable approach that could produce high‑quality, diverse images without these pitfalls. The field needed a stable, high‑quality generator.

The key idea

The key insight is to treat image generation as a denoising problem. First, an image is slowly corrupted into pure noise through many tiny Gaussian‑noise steps. Then a neural network is trained to learn the reverse: starting from noise, it removes the added noise step by step, eventually recovering a realistic image. This reframes generation as a sequence of simple denoising operations.

How it works

In the forward direction, the model adds a small amount of Gaussian noise at each of many timesteps, gradually turning any image into indistinguishable white noise. A U‑Net architecture is trained to predict the exact noise that was added at each step, effectively learning how to subtract it. During generation, we start with a random noise image and apply the trained network repeatedly, each time estimating and removing the noise for the current timestep. This reverse diffusion proceeds from the noisiest step back to the original image, with the network’s predictions guiding the process. The training objective is mathematically linked to score matching, ensuring the network learns accurate gradients of the data distribution.

See it — noise then denoise

flowchart LR
  A["Image"] --> B["Add noise<br/>step by step"]
  B --> C["Pure noise"]
  C --> D["Network removes<br/>noise step by step"]
  D --> E["New image"]

What they showed

The authors demonstrated that DDPMs achieved image quality that rivaled or even beat the best GANs at the time, while enjoying stable training and no mode collapse. Their results showed that diffusion models could match state‑of‑the‑art synthesis without the instability issues that plagued GANs.

What it changed

This paper shifted the research community’s focus toward diffusion models, making them the dominant approach for high‑quality image synthesis. By proving that diffusion could outperform GANs, it displaced GANs as the go‑to method for many state‑of‑the‑art generators. The success of DDPM sparked a wave of follow‑up work that refined and extended diffusion techniques across vision and multimodal tasks. Next: Latent Diffusion (Stable Diffusion).

The one-line takeaway

Diffusion models turn image generation into a stable, step‑by‑step denoising process that can outperform GANs.

Read it / meet it

Reading the original DDPM paper gives you a clear view of how the forward‑and‑reverse diffusion framework works and why it matters. Today, diffusion powers the most popular text‑to‑image systems, so understanding this work connects you directly to the technology you see in modern AI art tools.