Poisson Flow Generative Models

  • 2022-09-22 18:26:58
  • Yilun Xu, Ziming Liu, Max Tegmark, Tommi Jaakkola
  • 59

Abstract

We propose a new "Poisson flow" generative model (PFGM) that maps a uniformdistribution on a high-dimensional hemisphere into any data distribution. Weinterpret the data points as electrical charges on the $z=0$ hyperplane in aspace augmented with an additional dimension $z$, generating a high-dimensionalelectric field (the gradient of the solution to Poisson equation). We provethat if these charges flow upward along electric field lines, their initialdistribution in the $z=0$ plane transforms into a distribution on thehemisphere of radius $r$ that becomes uniform in the $r \to\infty$ limit. Tolearn the bijective transformation, we estimate the normalized field in theaugmented space. For sampling, we devise a backward ODE that is anchored by thephysically meaningful additional dimension: the samples hit the unaugmenteddata manifold when the $z$ reaches zero. Experimentally, PFGM achieves currentstate-of-the-art performance among the normalizing flow models on CIFAR-10,with an Inception score of $9.68$ and a FID score of $2.48$. It also performson par with the state-of-the-art SDE approaches while offering $10\times $ to$20 \times$ acceleration on image generation tasks. Additionally, PFGM appearsmore tolerant of estimation errors on a weaker network architecture and robustto the step size in the Euler method. The code is available athttps://github.com/Newbeeer/poisson_flow .

 

Quick Read (beta)

loading the full paper ...