The first diffusion model
Era 4 · Deep learning and foundation models (2012–2021) › Generative image models
- Lineage
- Neurons and networks
- Kind
- Publication
Who
Stanford University
What happened
Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli propose a generative model inspired by nonequilibrium thermodynamics: destroy the structure in the data little by little by adding noise, and learn to rebuild it in reverse.
Why it mattered
The first diffusion model. Five years later the same idea will become the foundation of the most widely used image generators.
The full story
The underlying problem with generative models was a trade-off: flexible models, able to represent complex data, were hard to train and evaluate, and the ones that were easy to handle were too simple.
Sohl-Dickstein and colleagues borrowed an idea from statistical physics. If noise is added to an image in many small steps, its structure is gradually erased until only pure noise remains, like a drop of ink spreading through water. That forward process is easy to describe. What the model learns is the way back: at each step, how to remove a little noise to recover some structure.
Because each step is small, each one can be learned with a simple model, and chaining thousands of them yields a very flexible generator that remains tractable: it can be trained, sampled from, and used to compute probabilities. The authors also released an open-source implementation.
For a few years the idea stayed in the shadow of adversarial networks. It took the work of Ho, Jain, and Abbeel in 2020 for diffusion to show everything it could do.
Leads to
DDPM (June 19, 2020)
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)
Threads that pass through here
- From noise to image (step 2 of 7)
Sources
Original source
- Deep Unsupervised Learning using Nonequilibrium Thermodynamics (arXiv:1503.03585) (opens in a new tab). Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli, 2015
Further reading
How to cite
“The first diffusion model.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/first-diffusion-model
This entry was last reviewed on November 13, 2026.