Large Scale GAN Training for High Fidelity Natural Image Synthesis

  • 2018-09-28 15:38:49
  • Andrew Brock, Jeff Donahue, Karen Simonyan
  • 409


Despite recent progress in generative image modeling, successfully generatinghigh-resolution, diverse samples from complex datasets such as ImageNet remainsan elusive goal. To this end, we train Generative Adversarial Networks at thelargest scale yet attempted, and study the instabilities specific to suchscale. We find that applying orthogonal regularization to the generator rendersit amenable to a simple "truncation trick", allowing fine control over thetrade-off between sample fidelity and variety by truncating the latent space.Our modifications lead to models which set the new state of the art inclass-conditional image synthesis. When trained on ImageNet at 128x128resolution, our models (BigGANs) achieve an Inception Score (IS) of 166.3 andFrechet Inception Distance (FID) of 9.6, improving over the previous best IS of52.52 and FID of 18.65.


Introduction (beta)



Conclusion (beta)