Diversity Augmented Conditional Generative Adversarial Network for Enhanced Multimodal Image-to-Image Translation

Conditional generative adversarial networks (cGANs) play an important role in multimodal image-to-image translation. We propose Diversity Augmented conditional Generative Adversarial Network (DivAugGAN), a highly effective solution to further resolve the mode collapse problem and enhance the diversity for the generated images. DivAugGAN functions as a regularizer to maximize the distinction of the generating samples when different noise vectors are injected. We also exert extra constraint on the generator to ensure the relative variation consistency in the translation process. This guarantees that the changing scale of the generated images in the image space is coherent to the difference of the injected noise vectors in the latent space. It also reduces the chances to bring about unexpected mode override and mode fusion issues. Experimental results on both two-domain and multi-domain multimodal image-to-image translation tasks demonstrate its effectiveness. DivAugGAN leads to consistent diversity augmentations and visual quality improvements for the developed models. We also achieves state-of-the-art performances on multiple datasets in terms of widely used quantitative evaluation metrics. DivAugGAN can be easily integrated into any objectives in conditional generative models as a regularizer for diversity augmentations and quality enhancements without any additional computation overheads compromise The source code and pre-trained models of our method is available at https://github.com/anomymous-gan/DivAugGAN.