Visual content is probably the most important medium by which we understand the world. In the digital age, millions of images and videos are uploaded to the Internet on each day. With the big amount of visual data and the rapid development of deep learning techniques, significant progress has been made in visual perception. However, visual synthesis, the opposite side of visual perception, is considered to be substantially more challenging as the model output space is much higher-dimensional and more complicated. The ability to automatically create visual content is not only important for various applications such as augmented reality, gaming, and photo editing, but also enables unsupervised representation learning that is central to building intelligent machines. In this dissertation, we present our work in developing better deep generative models that are capable of creating high-quality visual content. First, we introduce hierarchical unconditional generative models that learn to estimate the distribution of different types of visual content. We show how to model the distribution of 3D point clouds using hierarchical continuous normalizing flows, and how to model the distribution of images using hierarchical generative adversarial networks. The unconditional models are able to create visual content from scratch, which serve as the basis of conditional synthesis methods that aim to manipulate some given visual content. We then focus on conditional synthesis of images. We present adaptive instance normalization, an effective way to disentangle image content and style into normalized features and feature statistics respectively. We show how to apply adaptive instance normalization to manipulate the style of an image but preserve its content, while the style can be defined either by a single image or a collection of images.