DDPM
Era 4 · Deep learning and foundation models (2012–2021) › Generative image models
- Lineage
- Neurons and networks
- Kind
- Publication
Who
University of California, Berkeley
What happened
Jonathan Ho, Ajay Jain, and Pieter Abbeel of Berkeley present denoising diffusion probabilistic models (DDPM), which generate images of a quality comparable to that of the best adversarial networks.
Why it mattered
It turns diffusion into a practical, stable method for generating images. It is the basis of DALL·E 2, Imagen, Stable Diffusion, and the video generators that followed.
The full story
Five years after the Sohl-Dickstein paper, diffusion was still an elegant idea with modest results. Ho, Jain, and Abbeel found how to make it work.
The key was a connection to another line of work, the score-based models of Yang Song and Stefano Ermon, which learn to remove noise from data and generate samples with Langevin dynamics. By training the network to predict the noise that had been added to an image at each step, with a weighted version of the variational bound, they obtained their best results. Training is stable, without the delicate balance between two rival networks that GANs require.
On the CIFAR-10 image dataset the model reached an Inception score of 9.46 and an FID of 3.17, the best published for unconditional models. On 256-by-256-pixel LSUN images its quality was comparable to that of ProgressiveGAN, one of the best adversarial networks.
From then on diffusion gained ground quickly. In 2022 DALL·E 2, Imagen, and Stable Diffusion, the image generators that brought the technique to the general public, were all diffusion models, and in 2024 Sora applied the idea to video.
A row of boxes, from x0 (the image) to xT (pure noise). Above them, an arrow pointing right, "add noise." Below, a colored arrow pointing left, "remove noise," carried out by a neural network.
Under the hood
The forward process adds Gaussian noise over T = 1000 steps, with a variance βₜ that grows little by little, until nothing but noise is left of the image. It has a useful property: xₜ can be obtained directly from x₀ in a single calculation, without going through the intermediate steps.
The network, a U-Net, receives a noisy image and the step number, and predicts the noise that was added. The loss is the squared error between that noise and the real one, ‖ε − ε_θ(xₜ, t)‖². Ho, Jain, and Abbeel showed that this simplified form of the variational bound gave better samples than the full one.
To generate, the process starts from pure noise and applies the network a thousand times, removing a little noise at each step.
Side by side
| Feature | Variational autoencoder (2013) | Adversarial network (2014) | Diffusion (2020) |
|---|---|---|---|
| How it generates | Decodes a point drawn at random from the latent space | The generator turns noise into an image in one go | Removes noise from an image of pure noise over many steps |
| How it is trained | Maximizes a lower bound on the likelihood of the data | Through a game between generator and discriminator | Learns to predict the noise that was added |
| Steps to generate an image | One | One | Many (a thousand in this paper) |
Builds on
The first diffusion model (March 12, 2015)
Documented direct influence
DDPM defines diffusion models through the paper by Sohl-Dickstein et al., which it cites as their origin, and shows they can generate high-quality images (introduction)
Score-based generative models (July 12, 2019)
Documented direct influence
DDPM shows that one parameterization of diffusion models is equivalent to Song and Ermon's denoising score matching over multiple noise levels with annealed Langevin dynamics, and counts this among its main contributions (section 1)
Leads to
DALL·E 2 (April 6, 2022), Stable Diffusion (August 22, 2022)
Documented direct influence
The DALL·E 2 decoder is a diffusion model, and Stable Diffusion's latent diffusion applies diffusion models in a compressed space; both papers cite Ho et al. (2020)
Threads that pass through here
- From diffusion to world models (step 1 of 7)
- From noise to image (step 4 of 7)
Sources
Original source
- Denoising Diffusion Probabilistic Models (arXiv:2006.11239) (opens in a new tab). Jonathan Ho, Ajay Jain, and Pieter Abbeel, 2020
- Generative Modeling by Estimating Gradients of the Data Distribution (arXiv:1907.05600) (opens in a new tab). Yang Song and Stefano Ermon, 2019
Further reading
How to cite
“DDPM.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/ddpm
This entry was last reviewed on November 13, 2026.