Denoising Diffusion Probabilistic Models
Jonathan Ho, Ajay Jain, Pieter Abbeel
Summary
The paper introduces denoising diffusion probabilistic models (DDPMs), a class of latent-variable generative models trained to reverse a fixed Gaussian noising process. It establishes a connection between diffusion models and denoising score matching with Langevin dynamics, and proposes a simplified, reweighted training objective. The resulting models produce high-quality image samples, achieving competitive log-likelihoods and a strong FID on CIFAR-10.
Key findings
- Recasts diffusion model training as predicting the noise added at each step, yielding a simple weighted variational bound objective.
- Demonstrates an equivalence between diffusion models and score-based generative modeling with annealed Langevin sampling.
- Achieves state-of-the-art image generation quality at the time, including an FID of 3.17 on unconditional CIFAR-10.
Subjects & keywords
Cite this paper
Jonathan Ho, Ajay Jain, & Pieter Abbeel (2020). Denoising Diffusion Probabilistic Models. Advances in Neural Information Processing Systems 33 (NeurIPS 2020). https://arxiv.org/abs/2006.11239
@inproceedings{ho2020denoising,
author = {Jonathan Ho and Ajay Jain and Pieter Abbeel},
title = {Denoising Diffusion Probabilistic Models},
booktitle = {Advances in Neural Information Processing Systems 33 (NeurIPS 2020)},
year = {2020},
url = {https://arxiv.org/abs/2006.11239}
}