# Denoising Diffusion Probabilistic Models

**Authors:** Jonathan Ho, Ajay Jain, Pieter Abbeel  
**arXiv:** [2006.11239](https://arxiv.org/abs/2006.11239)  
**Date:** 2020-06-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2006.11239v2/

## Key points

- Authors present a novel approach to high-quality image synthesis using diffusion probabilistic models
- Models are inspired by concepts from nonequilibrium thermodynamics and belong to the class of latent variable models
- Impressive results achieved by training models on a weighted variational bound
- Unique connection between diffusion probabilistic models and denoising score matching with Langevin dynamics
- Introduction of a progressive lossy decompression scheme as a generalization of autoregressive decoding
- Performance evaluated on CIFAR10 and LSUN datasets, achieving high scores in Inception and FID metrics
- Sample quality comparable to ProgressiveGAN on LSUN dataset
- Code implementation provided on GitHub for further research and implementation

## Summary

In their paper titled "Denoising Diffusion Probabilistic Models," authors Jonathan Ho, Ajay Jain, and Pieter Abbeel present a novel approach to high-quality image synthesis using diffusion probabilistic models. These models are inspired by concepts from nonequilibrium thermodynamics and belong to the class of latent variable models. The authors achieve impressive results by training their models on a weighted variational bound. This bound is designed based on a unique connection between diffusion probabilistic models and denoising score matching with Langevin dynamics. Additionally, the authors introduce a progressive lossy decompression scheme that can be interpreted as a generalization of autoregressive decoding. The performance of the proposed models is evaluated on two datasets: unconditional CIFAR10 and 256x256 LSUN. On CIFAR10, the authors achieve an Inception score of 9.46 and a state-of-the-art FID score of 3.17. Moreover, on LSUN, their models demonstrate sample quality comparable to ProgressiveGAN. To facilitate further research and implementation, the authors provide their code implementation on GitHub at https://github.com/hojonathanho/diffusion. Overall, this paper introduces an innovative approach to image synthesis using diffusion probabilistic models. The combination of weighted variational bounds, denoising score matching with Langevin dynamics, and progressive lossy decompression contributes to achieving high-quality results on various datasets.