Optimizing GAN Training for Improved ImageDiversity and Quality

Since 2014, GANs have been extensively researched, leading to thedevelopment of numerous algorithms and techniques to improve GANs.Although modern techniques and algorithms for GANs are highly effec-tive, they tend to be very difficult to implement and take a long time totrain. Also, GANs are difficult to stabilize and cause problems such asmode collapse and suffer from the vanishing gradient problem, thus giv-ing poor results. This paper tries to improve the stability, quality, anddiversity of images generated by three GANs that are DCGANs ,CGANs,WGANs on datasets like MNIST, CIFAR10, FMNIST,and SVHN, usingmodern techniques that are easy to implement and take a short time toget results that are better than the basic implementation of these GANs.These techniques include label smoothing, minibatch standard deviation,and usage of reconstruction loss. The research done shows us that we canget good results in a short time using the techniques mentioned above.