Hierarchical Approach Towards High Fidelity Image Generation

Abstract

The high-fidelity image generation has been a subject of active research in the recent past. It provides benchmark towards image decoder’s performance. The autoregressive image models generate small images successfully but scalability has been a problem. The challenges include vast encoding of previous context and learning distribution which maintains global semantic coherence and exactness. These issues have been addressed through subscale pixel network (SPN) and multidimensional upscaling. To improve the accuracy further, in this work a hierarchical version of image generation model is presented. It disentangles background, object shape and appearance to hierarchically generate images of fine-grained object categories. To achieve this information theory associates a factor to latent code and condition relationships between codes to induce hierarchy. The hierarchical model’s learned features are used to cluster real images. The experimental results on ImageNet and CelebAHQ datasets for different image sizes highlight hierarchical model’s superiority against the benchmarks. The images are generated with better fidelity with respect to large scale samples.