14
65

PFGM++: Unlocking the Potential of Physics-Inspired Generative Models

Abstract

We introduce a new family of physics-inspired generative models termed PFGM++ that unifies diffusion models and Poisson Flow Generative Models (PFGM). These models realize generative trajectories for NN dimensional data by embedding paths in N+DN{+}D dimensional space while still controlling the progression with a simple scalar norm of the DD additional variables. The new models reduce to PFGM when D=1D{=}1 and to diffusion models when DD{\to}\infty. The flexibility of choosing DD allows us to trade off robustness against rigidity as increasing DD results in more concentrated coupling between the data and the additional variable norms. We dispense with the biased large batch field targets used in PFGM and instead provide an unbiased perturbation-based objective similar to diffusion models. To explore different choices of DD, we provide a direct alignment method for transferring well-tuned hyperparameters from diffusion models (DD{\to} \infty) to any finite DD values. Our experiments show that models with finite DD can be superior to previous state-of-the-art diffusion models on CIFAR-10/FFHQ 64×6464{\times}64 datasets, with FID scores of 1.91/2.431.91/2.43 when D=2048/128D{=}2048/128. In class-conditional setting, D=2048D{=}2048 yields current state-of-the-art FID of 1.741.74 on CIFAR-10. In addition, we demonstrate that models with smaller DD exhibit improved robustness against modeling errors. Code is available at https://github.com/Newbeeer/pfgmpp

View on arXiv
Comments on this paper