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

By Yilun Xu et al

Read the original document by opening this link in a new tab.

Table of Contents

1. Introduction
2. Background and Related Works
3. PFGM++: A Novel Generative Framework
4. Diffusion Models as D!1 Special Cases

Summary

PFGM++ introduces a new family of physics-inspired generative models that unify diffusion models and Poisson Flow Generative Models. It extends the electrostatic view into higher dimensions through multi-dimensional augmentations. The models provide generative paths based on the scalar norm of augmented variables, offering a balance between robustness and rigidity controlled by the augmentation dimension D. A perturbation-based objective is proposed to train the models efficiently without relying on large batches. The document also discusses how PFGM++ can be seen as a special case of diffusion models when D approaches 1.