This work introduces a novel approach linking Stochastic Partial Differential Equations (SPDEs) to Score-based Generative Models (SGMs) via the stochastic Fokker-Planck equation. This connection aims to provide a deeper understanding of the underlyin...
This work introduces a novel approach linking Stochastic Partial Differential Equations (SPDEs) to Score-based Generative Models (SGMs) via the stochastic Fokker-Planck equation. This connection aims to provide a deeper understanding of the underlying dynamics in generative modeling, reflecting substantive, real-world phenomena. We first derive an error function D(t,x) from the Fokker-Planck equation of the SGM's reverse process, representing the time-varying difference between model and target distributions. We then assume that this error function D(t,x) is a sample path of an infinite-dimensional stochastic process Κ(t,x). We design a new forward SPDE, a stochastic Fokker-Planck type equation, fundamentally driven by this Κ(t,x). We prove the existence and uniqueness of the SPDE solution, ensuring our analytical framework is well-grounded. This SPDE formulation allows us to quantify a discrepancy α between its solution (driven by Κ(t,x)) and a hypothetical trajectory without this driving term. For computational tractability, we assume the Wiener process within Κ(t,x) to be spatially homogeneous. Since the covariance of such a process is related to the Fourier transform, utilizing Fourier transform techniques reduces the computation of the discrepancy α, originally an infinite-dimensional problem, to a tractable, single one-dimensional problem. We then collect samples for the computation of this one-dimensional problem utilizing the ergodic property. We also employ a technique similar to denoising score matching, allowing for a straightforward implementation akin to SGM implementations. This analysis yields two scalar terms, Trend and Residuals, satisfying α^2 \leq K(\text{Trend}^2 +\text{Residuals}^2) for some K>0. We call the pair of Trend and Residuals the SPDE-Induced Evaluation Metric (SIEM). Interestingly, the Residuals component correlates with the other evaluation metrics during model training. This correlation, rooted in our SPDE formulation, validates that our mathematical framework captures meaningful dynamics of generative model quality, thereby demonstrating its empirical grounding. Furthermore, we show that SIEM achieves this reliability with significantly fewer sampling steps, highlighting its potential for substantial computational efficiency. Finally, we conclude by highlighting the potential of our SPDE framework to offer new insights into model discrepancies and suggesting avenues for future research.