Title: Accelerating Electrostatic Generative Models via Overclocking

Abstract: Electrostatic generative frameworks, notably PFGM++, have recently established themselves as robust tools for image synthesis, delivering competitive results. These models function within an extended data space characterized by an auxiliary dimensionality $D$. As $D$ approaches infinity, the PFGM++ framework converges to the standard diffusion model architecture, yet it demonstrates empirically superior performance when $D$ remains finite. Similar to diffusion models, PFGM++ generation processes depend on computationally intensive ODE simulations. To mitigate this high cost, we introduce Inverse Poisson Flow Matching (IPFM), a systematic distillation method designed to speed up electrostatic generative models regardless of the value of $D$. IPFM reimagines distillation as an inverse problem, aiming to train a generator that produces an electrostatic field identical to that of the teacher model. We develop a feasible training objective for this approach and demonstrate that in the limit where $D\to\infty$, IPFM effectively replicates Score Identity Distillation (SiD), a contemporary technique for distilling diffusion models. Our empirical findings indicate that generators distilled via IPFM attain sample quality comparable to or exceeding that of the teacher, requiring only a minimal number of function evaluations. Furthermore, we observe that distilling one-step generators converges more rapidly at finite $D$ than in the infinite $D$ diffusion limit. This observation supports previous evidence suggesting that finite-$D$ PFGM++ models provide more advantageous dynamics for both optimization and sampling.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC