MSG-GAN: Multi-Scale Gradients GAN for more stable and synchronized multi-scale image synthesis

  • 2019-03-14 14:33:26
  • Animesh Karnewar, Raghu Sesha Iyengar
  • 12

Abstract

Generative Adversarial Network (GAN) which is widely used for Image synthesisvia generative modelling suffers peculiarly from training instability. One ofthe known reasons for this instability is the passage of uninformativegradients from the Discriminator to the Generator due to learning imbalancebetween them during training. In this work, we propose Multi-Scale GradientsGenerative Adversarial Network (MSG-GAN), a simplistic but effective techniquefor addressing this problem; by allowing the flow of gradients from theDiscriminator to the Generator at multiple scales. This results in theGenerator acquiring the ability to synthesize synchronized images at multipleresolutions simultaneously. We also highlight a suite of techniques thattogether buttress the stability of training without excessive hyperparametertuning. Our MSG-GAN technique is a generic mathematical framework which hasmultiple instantiations. We present an intuitive form of this technique whichuses the concatenation operation in the Discriminator computations andempirically validate it through experiments on the CelebA-HQ, CIFAR10 andOxford102 flowers datasets and by comparing it with some of the currentstate-of-the-art techniques.

 

Introduction (beta)

None

 

Conclusion (beta)

None