The History of AI

Score-based generative models

Era 4 · Deep learning and foundation models (2012–2021) › Generative image models

Lineage
Neurons and networks
Kind
Publication

Who

Stanford University

What happened

Yang Song and Stefano Ermon of Stanford post on arXiv a model that generates images by starting from noise and following, step by step, the direction in which the data become more probable. It is presented as an oral paper at NeurIPS 2019.

Why it mattered

It shows that learning to remove noise at several levels is enough to generate images comparable in quality to those of GANs. DDPM will join this idea with diffusion in 2020.

The full story

In 2019 GANs dominated image generation, with their delicate balance between two networks. Song and Ermon looked for another route: learning the score of the data, that is, the gradient that shows in which direction to change an image to make it look more like real images.

Estimating that gradient is hard where there are almost no data, so they estimated it on versions of the data perturbed with Gaussian noise at several levels, using a single network for all of them. To generate, they started from noise and applied annealed Langevin dynamics: moving along the estimated score with ever smaller noise levels as the sample approaches the real data.

The method needed neither adversarial training nor sampling during training. On MNIST, CelebA, and CIFAR-10 it produced images comparable to those of GANs, with an Inception score of 8.87 on CIFAR-10, the best published for that dataset up to then.

A year later Ho, Jain, and Abbeel showed that one way of parameterizing diffusion models is equivalent to this multilevel denoising with Langevin dynamics, and they counted it among their main contributions. The two lines merged in the image generators that followed.

Leads to

  • DDPM (June 19, 2020)

    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)

Threads that pass through here

Sources

How to cite

“Score-based generative models.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/score-based-generative-models

This entry was last reviewed on November 13, 2026.

Open in the interactive timeline