Reversible GANs for Memory-efficient Image-to-Image Translation

  • 2019-02-07 17:03:05
  • Tycho F. A. van der Ouderaa, Daniel E. Worrall
  • 24

Abstract

The Pix2pix and CycleGAN losses have vastly improved the qualitative andquantitative visual quality of results in image-to-image translation tasks. Weextend this framework by exploring approximately invertible architectures whichare well suited to these losses. These architectures are approximatelyinvertible by design and thus partially satisfy cycle-consistency beforetraining even begins. Furthermore, since invertible architectures have constantmemory complexity in depth, these models can be built arbitrarily deep. We areable to demonstrate superior quantitative output on the Cityscapes and Mapsdatasets at near constant memory budget.

 

Introduction (beta)

None

 

Conclusion (beta)

None