Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Mirror Diffusion Models for Constrained and Watermarked Generation

About

Modern successes of diffusion models in learning complex, high-dimensional data distributions are attributed, in part, to their capability to construct diffusion processes with analytic transition kernels and score functions. The tractability results in a simulation-free framework with stable regression losses, from which reversed, generative processes can be learned at scale. However, when data is confined to a constrained set as opposed to a standard Euclidean space, these desirable characteristics appear to be lost based on prior attempts. In this work, we propose Mirror Diffusion Models (MDM), a new class of diffusion models that generate data on convex constrained sets without losing any tractability. This is achieved by learning diffusion processes in a dual space constructed from a mirror map, which, crucially, is a standard Euclidean space. We derive efficient computation of mirror maps for popular constrained sets, such as simplices and $\ell_2$-balls, showing significantly improved performance of MDM over existing methods. For safety and privacy purposes, we also explore constrained sets as a new mechanism to embed invisible but quantitative information (i.e., watermarks) in generated data, for which MDM serves as a compelling approach. Our work brings new algorithmic opportunities for learning tractable diffusion on complex domains. Our code is available at https://github.com/ghliu/mdm

Guan-Horng Liu, Tianrong Chen, Evangelos A. Theodorou, Molei Tao• 2023

Related benchmarks

TaskDatasetResultRank
Constrained GenerationSimplices (test)
W1 Distance1
15
Unconditional Watermarked GenerationFFHQ 64x64 (train)
FID (50k)2.54
7
Unconditional Watermarked GenerationAFHQ 64x64 v2 (train)
FID (50k Samples)2.1
7
Constrained Generative ModelingHypercube [0, 1]^d d=3 (test)
Sliced Wasserstein0.0192
4
Constrained Generative ModelingHypercube [0, 1]^d d=6 (test)
SWD1.75
4
Constrained Generative ModelingHypercube [0, 1]^d d=8 (test)
Sliced Wasserstein Distance0.0185
4
Constrained Generative ModelingHypercube [0, 1]^d d=2 (test)
Sliced Wasserstein Distance0.03
4
Constrained Generative ModelingHypercube [0, 1]^d d=20 (test)
SWD0.0335
4
Constrained Generationl2-ball constrained set (d=2, Gaussian Mixture) 1.0 (synthetic)
Constraint Violation Rate0.00e+0
2
Constrained Generationl2-ball constrained set d=2 Spiral 1.0 (synthetic)
Constraint Violation Rate0.00e+0
2
Showing 10 of 13 rows

Other info

Follow for update