Atlas record: https://www.aitimeline.dev/en/events/evt-0485/
Research · June 19, 2020
Denoising diffusion probabilistic models
A denoising diffusion model reached image quality that had belonged to adversarial networks.
Why it matters
Diffusion stopped being a described route to image synthesis and became a working one. The atlas holds the 2015 paper that described the process; this is the moment it produced numbers comparable to the best adversarial networks, and the generators that followed take this form.
The paper reports an Inception score of 9.46 and an FID of 3.17 on unconditional CIFAR-10, and sample quality on 256x256 LSUN comparable to ProgressiveGAN. The authors are Jonathan Ho, Ajay Jain and Pieter Abbeel of UC Berkeley.
It is a result on those datasets, not a claim about every later image or video generator.
Event record
- Event date
- June 19, 2020
- Timeline date
- Event date
- Verification
- Sources gathered automatically · September 21, 2026
- Lines
- ID
- evt-0485
The day the first version of the preprint appeared on arXiv.
Related events
Earlier
- Enables U-Net: segmentation from a few images 2015links to this record
The DDPM network is, by its own description, a U-Net based on a Wide ResNet as in PixelCNN++, citing this paper.
- Builds on Diffusion models 2015
The same forward and reverse diffusion described in 2015; this paper builds training on a weighted variational bound on top of it.
- Contrasts Generative adversarial networks 2014
A different route to the same task: sample quality is measured against adversarial networks, without adversarial training.
- Enables CIFAR-10: tiny images with labels 2009links to this record
Diffusion models show their quality on unconditional CIFAR10 generation.
Later