# 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/2006.11239v1/

## Key points

- Authors: Jonathan Ho, Ajay Jain, Pieter Abbeel
- Introduction of novel approach to image synthesis using diffusion probabilistic models
- Inspired by concepts from nonequilibrium thermodynamics
- High-quality results through weighted variational bound training method
- Connection between diffusion probabilistic models and denoising score matching with Langevin dynamics
- Incorporation of progressive lossy decompression scheme as a generalization of autoregressive decoding
- Efficient image synthesis and generation demonstrated on CIFAR10 dataset
- Impressive Inception score of 9.46 and FID score of 3.17 achieved on CIFAR10 dataset
- Comparable sample quality to ProgressiveGAN on challenging 256x256 LSUN dataset
- Implementation available on GitHub at https://github.com/hojonathanho/diffusion

## Summary

In their paper titled "Denoising Diffusion Probabilistic Models," authors Jonathan Ho, Ajay Jain, and Pieter Abbeel introduce a novel approach to image synthesis using diffusion probabilistic models. These models are inspired by concepts from nonequilibrium thermodynamics and offer high-quality results through a weighted variational bound training method. By establishing a unique connection between diffusion probabilistic models and denoising score matching with Langevin dynamics, the authors achieve impressive outcomes. One key feature of their approach is the incorporation of a progressive lossy decompression scheme, which serves as a generalization of autoregressive decoding. This innovative technique allows for efficient image synthesis and generation. The authors demonstrate the effectiveness of their method on the CIFAR10 dataset, achieving an impressive Inception score of 9.46 and a state-of-the-art FID score of 3.17. Furthermore, when applied to the challenging 256x256 LSUN dataset, their model produces sample quality comparable to ProgressiveGAN. The implementation of their approach is openly available on GitHub at https://github.com/hojonathanho/diffusion, enabling further exploration and experimentation in the field of image synthesis. Overall, this work showcases the potential of diffusion probabilistic models in generating high-quality images and opens up new possibilities for future research in this area.