Latent Nonlinear Denoising Score Matching for Enhanced Learning of Structured Distributions

TL;DR

This paper introduces LNDSM, a new score-based generative model training method that combines nonlinear dynamics with VAEs, leading to faster and more accurate learning of structured data distributions.

Contribution

The paper proposes LNDSM, integrating nonlinear forward dynamics into latent score models using a reformulated cross-entropy and stability improvements, advancing structured distribution learning.

Findings

Faster synthesis of structured distributions.

Superior sample quality and variability.

Enhanced learning on MNIST variants.

Abstract

We present latent nonlinear denoising score matching (LNDSM), a novel training objective for score-based generative models that integrates nonlinear forward dynamics with the VAE-based latent SGM framework. This combination is achieved by reformulating the cross-entropy term using the approximate Gaussian transition induced by the Euler-Maruyama scheme. To ensure numerical stability, we identify and remove two zero-mean but variance exploding terms arising from small time steps. Experiments on variants of the MNIST dataset demonstrate that the proposed method achieves faster synthesis and enhanced learning of inherently structured distributions. Compared to benchmark structure-agnostic latent SGMs, LNDSM consistently attains superior sample quality and variability.