Disentangling Image Features using Laplacian Pyramid in Hierarchical Vector Quantized-Variational Autoencoder Engineering Pohang University of Science and Technology

Abstract

We present a novel approach for high-resolution image generation within the framework of Vector Quantized Variational Autoencoder (VQ-VAE). Building upon the advancements of VQ-VAE-2, which introduces a hierarchical architecture design to address challenges such as the need for larger codebooks and feature disentanglement, we extend this approach by incorporating Laplacian pyramid-based image decomposition. This extension allows for frequency-specific supervision, resulting in a more faithful representation of the underlying data and higher-quality generated images. Our experimental results confirm the superiority of our approach in terms of feature disentanglement and output fidelity, establishing it as a promising avenue for achieving high-quality image generation.