Holographic displays provide an immersive visualization of three-dimensional scenes by reconstructing light wavefronts that preserve both amplitude and phase information. Computer-generated holography enables the numerical creation of holograms without complex optical setups, offering scalability for 3D content generation. However, conventional Computer-Generated Holography techniques face critical challenges in balancing computational cost, storage efficiency, and visual quality, hindering their use in real-time and resource-constrained applications. The first method employs a Haar wavelet decomposition to generate multilevel hologram representations. By incorporating object saliency, the approach emphasizes perceptually important regions while significantly reducing computation time during offline dataset generation. The second method explores the capability of Vector Quantized Variational Autoencoder with hierarchical latent spaces to effectively encode complex-valued holograms, achieving high-fidelity compression through training on the diverse InterfereI dataset. This framework provides a compact and robust hologram representation suitable for holographic applications. Building upon these advancements, the third contribution introduces our Rate Adaptive Vector Quantized Variational Autoencoder framework, designed to achieve flexible compression within a single network. Our proposed network delivers high-quality reconstructions in real-time for phase-only hologram at both low and ultra-low bit rates, outperforming a state-of-the-art method on the natural image dataset of Div2K with a BD-Rate reduction of $-33.91% and a BD-PSNR improvement of 1.02 dB. The proposed method paves the way toward practical and scalable holographic systems for next-generation 3D display technologies, such as virtual reality near-eye display.