Hierarchy Composition GAN for High-fidelity Image Synthesis

Abstract

Despite the rapid progress of generative adversarial networks (GANs) in image synthesis in recent years, the existing image synthesis approaches work in either geometry domain or appearance domain alone which often introduces various synthesis artifacts. This paper presents an innovative Hierarchical Composition GAN (HIC-GAN) that incorporates image synthesis in geometry and appearance domains into an end-to-end trainable network and achieves superior synthesis realism in both domains simultaneously. We design an innovative hierarchical composition mechanism that is capable of learning realistic composition geometry and handling occlusions while multiple foreground objects are involved in image composition. In addition, we introduce a novel attention mask mechanism that guides to adapt the appearance of foreground objects which also helps to provide better training reference for learning in geometry domain. Extensive experiments on scene text image synthesis, portrait editing and indoor rendering tasks show that the proposed HIC-GAN achieves superior synthesis performance qualitatively and quantitatively.

This work has been merged into another project

References in corpus (8)

• GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium
• Toward Multimodal Image-to-Image Translation
• Synthetic Data and Artificial Neural Networks for Natural Scene Text Recognition
• LR-GAN: Layered Recursive Generative Adversarial Networks for Image Generation
• Scene Text Synthesis for Efficient and Effective Deep Network Training
• Towards Realistic 3D Embedding via View Alignment
• Fast Guided Filter
• EMLight: Lighting Estimation via Spherical Distribution Approximation