Class-Splitting Generative Adversarial Networks
arXiv:1709.07359
Abstract
Generative Adversarial Networks (GANs) produce systematically better quality samples when class label information is provided., i.e. in the conditional GAN setup. This is still observed for the recently proposed Wasserstein GAN formulation which stabilized adversarial training and allows considering high capacity network architectures such as ResNet. In this work we show how to boost conditional GAN by augmenting available class labels. The new classes come from clustering in the representation space learned by the same GAN model. The proposed strategy is also feasible when no class information is available, i.e. in the unsupervised setup. Our generated samples reach state-of-the-art Inception scores for CIFAR-10 and STL-10 datasets in both supervised and unsupervised setup.
Under consideration at Pattern Recognition Letters
References in corpus (3)
Cited by in corpus (17)
- Progressive Growing of GANs for Improved Quality, Stability, and Variation
- TransGAN: Two Pure Transformers Can Make One Strong GAN, and That Can Scale Up
- Deep Polynomial Neural Networks
- Activation Maximization Generative Adversarial Nets
- AutoGAN: Neural Architecture Search for Generative Adversarial Networks
- AGAN: Towards Automated Design of Generative Adversarial Networks
- Transferring GANs: generating images from limited data
- AdversarialNAS: Adversarial Neural Architecture Search for GANs
- cGANs with Multi-Hinge Loss
- Generative Feature Replay For Class-Incremental Learning
- JGAN: A Joint Formulation of GAN for Synthesizing Images and Labels
- Score-Guided Generative Adversarial Networks
- GAN-Based Facial Attractiveness Enhancement
- Lessons Learned from the Training of GANs on Artificial Datasets
- Virtual Conditional Generative Adversarial Networks
- Conditional Transferring Features: Scaling GANs to Thousands of Classes with 30% Less High-quality Data for Training
- Instance-Conditioned GAN