GPM: Unifying GANs and Score-Based Diffusion as Generative Particle Models

This page is dedicated to hosting errata and animated samples for the paper “Unifying GANs and Score-Based Diffusion as Generative Particle Models” (Jean-Yves Franceschi, Mike Gartrell, Ludovic Dos Santos, Thibaut Issenhuth, Emmanuel de Bézenac, Mickaël Chen, Alain Rakotomamonjy), to appear at NeurIPS 2023. Cf. the paper and the code for more information.

Errata

The careful reader may have noticed the following minor inconsistencies in Section 2.2 of the published paper. These small oversights do not affect the overall conclusions of the article.

  • All references to the time derivative of \sigma should be understood in absolute value: \sigma' \equiv |\sigma'|.
  • The Gaussian kernel in Eq. (5) should be read as:

    $$k_{\mathrm{RBF}}^\sigma (x, y) \triangleq \frac{1}{(\sigma \sqrt{2\pi})^{D}} \mathrm{e}^{-\frac{\|x-y\|_2^2}{2\sigma^2}}.$$

Animated Samples

MNIST

EDM

Discriminator Flow

CelebA

EDM

Discriminator Flow

Gaussians

Please refer to Appendix C of the paper for a detailed description.

EDM

55 NFE
13 NFE

Discriminator Flow (inference)

56 NFE
14 NFE

GAN (training)

Score GAN (training)

Generated samples.
Noisy samples with corresponding gradients.