Skip to main content

Research · Technical

Denoising Diffusion Probabilistic Models (DDPM)

NeurIPS 2020 · Ho, J., Jain, A., Abbeel, P. · Jun 2020

Key points

  1. DDPM learns to reverse a fixed noising process, generating images by iterative denoising.
  2. It is the direct mathematical foundation of latent diffusion, AnimateDiff, Video LDM and MDM.
  3. It is taught as 'the diffusion model' in ML-for-creatives courses.

Summary

Ho, Jain and Abbeel (2020) demonstrated that a neural network trained to reverse a fixed Markov noising chain can generate high-quality images by iteratively denoising Gaussian noise. The denoising score-matching objective gives DDPM a stable training regime compared to GANs, and the resulting sample quality matched or exceeded the state of the art. DDPM is the pixel-space origin of the entire diffusion stack: Stable Diffusion, AnimateDiff, Video LDM and the Motion Diffusion Model all build on its mathematical foundation. It is the paper courses name when introducing diffusion models to creative-AI students.

Source

Source: NeurIPS 2020 ↗ (Research)

Cite this item

Ho, J., Jain, A. & Abbeel, P. (2020). ‘Denoising Diffusion Probabilistic Models (DDPM)’, NeurIPS 2020. Available at: https://arxiv.org/abs/2006.11239

Your reference manager can also read this page directly: with the Zotero (or Mendeley) browser connector installed, save it straight to your library. Whole-collection exports: RIS, BibTeX, CSL-JSON.