Skip to content

This repository investigates the relationship between DDPMs and OT theory, as outlined in the paper "Understanding DDPM Latent Codes through Optimal Transport" by Khrulkov (2022).

Notifications You must be signed in to change notification settings

lucasdegeorge/diffusion_ot

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

25 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Denoising Diffusion Probabilistic Models (DDPMs) and Optimal Transport (OT)

This repository explores the intersection between DDPMs and OT theory.

DDPMs are generative models that utilize a forward diffusion process to transform clean data into Gaussian distributions by gradually adding noise. The paper titled Understanding DDPM Latent Codes through Optimal Transport (Khrulkov 2022) provides insights into this aspect by exploring the relationship between DDPM latent codes and optimal transport theory.

Theory

The Monge map seeks to minimize transportation costs between probability distributions, while the DDPM encoder map aims to map data distributions to a standard normal distribution. Theoretical results establish that the DDPM encoder map achieves the minimum transportation cost, aligning with the Monge optimal transport map.

Example Code

The repository contains code snippets and examples demonstrating the theoretical concepts discussed above. It includes implementations of DDPMs and OT computations using libraries such as JAX, OTT JAX, and PyTorch. For detailed explanations, please refer to the corresponding sections and files within the repository.

About

This repository investigates the relationship between DDPMs and OT theory, as outlined in the paper "Understanding DDPM Latent Codes through Optimal Transport" by Khrulkov (2022).

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published