Paper breakdown

DDPM

The 2020 paper that proved diffusion models could generate high-quality images by learning to reverse a gradual noising process.

Paper: Denoising Diffusion Probabilistic Models

Authors: Jonathan Ho, Ajay Jain, Pieter Abbeel · 2020

Read the paper

The forward process gradually destroys an image by adding Gaussian noise until it becomes pure static. The reverse process (the neural network) learns to denoise it step by step.

The Problem

Before 2020, Generative Adversarial Networks (GANs) ruled image generation. While GANs produced sharp images, they were notoriously unstable to train and suffered from "mode collapse" (only generating a few types of images). Variational Autoencoders (VAEs) and Normalizing Flows were stable and mathematically sound, but their images were blurry or lacked fine detail. The field needed a model that combined the stability and diversity of likelihood-based models with the high-quality samples of GANs.

The Idea

The authors built upon the concept of diffusion models (first proposed in 2015). The idea is rooted in non-equilibrium thermodynamics. If you add noise to an image step-by-step, it eventually becomes pure static (a Gaussian distribution). This "forward process" is easy. The breakthrough was showing that a neural network could learn the exact "reverse process": starting from pure noise, the network predicts and subtracts a tiny amount of noise at each step, eventually revealing a pristine, generated image.

How It Works

DDPM operates in two chains:

Forward Process (No Training): Starting with a real image $x_0$ , a fixed schedule of Gaussian noise is added over $T$ steps (e.g., $T=1000$ ). At step $x_T$ , the image is indistinguishable from pure noise.
Reverse Process (The Neural Network): A U-Net neural network is trained to look at a noisy image $x_t$ and the current timestep $t$ , and predict the noise that was added at that specific step.

The training objective is remarkably simple: it is just the Mean Squared Error (MSE) between the actual noise added during the forward process and the noise predicted by the U-Net. To generate a new image, you sample pure noise $x_T$ , and run it through the trained U-Net $T$ times, subtracting the predicted noise at each step.

Why It Mattered

DDPM was the "AlexNet moment" for generative AI. It proved that diffusion models could match or beat GANs in image quality without mode collapse or adversarial training instability. It laid the foundation for the entire modern explosion of generative image and video models.

What Came After

While DDPM proved the concept, generating an image required 1,000 passes through a massive U-Net, making it glacially slow. The immediate next wave of research (DDIM, Latent Diffusion) focused entirely on speeding up this sampling process, leading to the models that power DALL-E, Stable Diffusion, and Midjourney today.