Generating high-quality and various image samples is a significant research goal in computer vision and artificial intelligence. The Generative Adversarial Networks (GAN) and Variational Autoencoder (VAE) are widely used to capture the distributions of actual distributions samples. In this thesis, we pay attention to the GANs, a prominent unsupervised learning method that can automatically capture the patterns in the training data. However, the training of GANs has simple memory imitation and non-convergence issues. The memory imitation issue means generators produce same samples lazily to fool discriminators. To generate various image samples and improve the GANs’ performance efficiently, we develop the GAN structure from the following three aspects: 1) The training procedure is not stable enough, which incurs the mode collapse issues. The mode collapse means the GAN will generate samples with single diversity; 2) The training process requires enormous time to capture the pattern from the training data. The complexity of GAN structure and the amount of training data influence the total training expense; 3) GAN demands enough training data to ensure the accuracy and stability of the model. Lack of comprehensive training data usually causes deterioration of the performance of the network. Thus, we investigate training techniques and propose the framework to develop the GANs’ performance and ability. First, we present the Multi-group GAN (MGGAN), a light framework to solve the mode collapse while increasing the diversity of generated samples. Next, we present the Block Paralleling GAN (BPGAN) to decrease the total training time. It uses a novel model parallelism to reduce the transmission cost. We provide theoretical analysis to prove the benefit of our method. Finally, we present Privacy-aware GAN (PrivacyGAN), a teacher-student framework based on a generative adversarial network, to generate similar sensitive personal data from private clients. The experimental results and theoretical analysis demonstrate that the techniques proposed in this thesis is effective.