Banach Wasserstein GAN

Jonas Adler, Sebastian Lunz
6/18/2018
cs.CVcs.LGmath.FA

Official code: adler-j/bwgan (31 stars, last commit 2019-03-10).

Abstract

Wasserstein Generative Adversarial Networks (WGANs) can be used to generate realistic samples from complicated image distributions. The Wasserstein metric used in WGANs is based on a notion of distance between individual images, which induces a notion of distance between probability distributions of images. So far the community has considered $\ell^2$ as the underlying distance. We generalize the theory of WGAN with gradient penalty to Banach spaces, allowing practitioners to select the features to emphasize in the generator. We further discuss the effect of some particular choices of underlying norms, focusing on Sobolev norms. Finally, we demonstrate a boost in performance for an appropriate choice of norm on CIFAR-10 and CelebA.

View on arXivView PDF

Code Implementations(1)

adler-j/bwganOfficial100%

Code for the paper "Banach Wasserstein GAN"

3111PythonMay 23, 20187 years ago

Similar Papers

Rethinking FID: Towards a Better Evaluation Metric for Image Generation

83% match1 repo39.0k2024

Wasserstein Geodesic Generator for Conditional Distributions

82% match1 repo32023

Learning a distance measure from the information-estimation geometry of data

82% match1 repo132025

Generalized Probabilistic U-Net for medical image segementation

82% match1 repo162022

A Geometric Unification of Generative AI with Manifold-Probabilistic Projection Models

81% match2025

Fréchet Wavelet Distance: A Domain-Agnostic Metric for Image Generation

81% match1 repo182023

Cite this paper

@article{adler2018banach,
  title  = {Banach Wasserstein GAN},
  author = {Jonas Adler and Sebastian Lunz},
  year   = {2018},
  eprint = {1806.06621},
  archivePrefix = {arXiv},
  url    = {https://arxiv.org/abs/1806.06621}
}

Add a badge to your README

Shows whether official code is linked for this paper.

Code of Paper badge
[![Code of Paper](https://codeofpaper.com/badge/1806.06621.svg)](https://codeofpaper.com/paper/1806.06621)

Discussion