Hierarchically-Fused Generative Adversarial Network for Text to Realistic Image Synthesis

In this paper, we present a novel Hierarchically-fused Generative Adversarial Network (HfGAN) for synthesizing realistic images from text descriptions. While existing approaches on this topic have achieved impressive success, to generate 256X256 images from captions, they commonly resort to coarse-to-fine scheme and associate multiple discriminators in different stages of the networks. Such a strategy is both inefficient and prone to artifacts. Motivated by the above findings, we propose an end-to-end network that can generate 256X256 photo-realistic images with only one discriminator. We fully exploit the hierarchical information from different layers and directly generate the fine-scale images by adaptively fusing features from multi-hierarchical layers. We quantitatively evaluate the synthesized images with Inception Score, Visual-semantic Similarity and average training time on the CUB birds, Oxford-102 flowers, and COCO datasets. The results show that our model is more efficient and noticeably outperforms the previous state-of-the-art methods.