CAGAN: Text-To-Image Generation with Combined Attention Generative Adversarial Networks

Generating images according to natural language descriptions is a challenging task. Prior research has mainly focused enhance the quality of generation by investigating use spatial attention and/or textual thereby neglecting relationship between channels. In this work, we propose Combined Attention Generative Adversarial Network (CAGAN) generate photo-realistic descriptions. The proposed CAGAN utilises two models: word draw different sub-regions conditioned on related words; and squeeze-and-excitation capture non-linear interaction among With spectral normalisation stabilise training, our achieves state-of-the-art FID comparative IS scores CUB dataset more COCO dataset. Furthermore, demonstrate that judging model single evaluation metric can be misleading developing an additional adding local self-attention which higher than other model, but generates unrealistic through feature repetition.