This thesis exploits deep learning to develop real-time, super-resolution hologram generation. It first gives an overview of both the technologies of deep learning and holographic 3D displays. It demonstrates the strength of learning models and also identifies the challenges and limitations of conventional hologram computation methods for 3D scenes. These motivate us to explore the issues and seek solutions via deep generative models. We propose an original concept of introducing frequency modulators to allow the use of generative models to interpret frequency data directly. After training, this new mechanism enables the learning model to generate frequency samples with variations in the underlying generative features. Furthermore, a new generative model is proposed, named the channelled variational autoencoder (CVAE). We design this model specifically for hologram generation. Similar to generating frequency data, we use the spatial spectra of the hologram modulators (SSHM) as the training dataset. The trained CVAE can then interpret and learn the hidden structure of input holograms. It is thus able to generate holograms through the learning of the disentangled latent representations. For practical purposes, we propose a new technique called hologram super-resolution (HSR). It exploits a supervised learning model to function SSHM inputs to SSHM outputs. As a result, a well-trained HSR model is capable of super-resolving a low-resolution hologram input to a super-resolution hologram output. Based on the joint model of the proposed CVAE and HSR, we finally develop a new approach of producing super-resolved, complex-amplitude holograms for 3D scenes in real-time. It takes a 3D object as a basic element and then adds the holograms of all the 3D objects of one 3D scene together to obtain the ultimate hologram. The superposed hologram is able to reconstruct entire 3D objects. Most notably, this object-based method is much faster than traditional ways. Finally, the research examines the replay image quality of the generated hologram produced by the proposed approaches in two major holographic display systems. As the results show, it successfully demonstrates adequate holographic 3D displays.