In recent years, there have been huge success of supervised learning in computer vision applications. Meanwhile unsupervised representation learning has received attention especially for the generative adversarial networks (GANs) for images generation. In the report, we adopt recent advancements in GAN, Wasserstein GAN and deep convolutional generative adversarial network (DCGAN) to generate sky images. We take rigorous experiments on LSUN dataset and our sky images dataset scrawled from the Internet. Our experiments show hierarchy of representations learning ability from object parts to scenes in both the generator and discriminator of DCGAN and Wasserstein GAN’s theory validity to address common problems during training process of original DCGAN.