Piecewise Diffusive Score-based Generative Model

Abstract

Diffusion-based generative models have achieved remarkable success across a range of applications. However, their performance often degrades in the presence of imbalanced data. To address this limitation, we introduce a novel generative modeling framework that combines stochastic differential equation (SDE) driven by Brownian noise with Poisson random measures. We also provide the explicit form of reverse SDE, with corresponding loss function to train the neural network. Our experimental results show that the model with Poisson jump outperforms the model without jump across various toy datasets and a small subset of CIFAR-10.