Two-stage effective attentional generative adversarial network

Mingyu Jin, Qinkai Yu, Chong Zhang, Haochen Xue, Shuliang Zhao

Abstract: Although GAN models have succeeded in relevant tests, text-to-image modelling using GANs to synthesize high-quality images is still challenging. Existing multi-stage models face several problems: first, the scale is too large, and the model has a large number of redundant structures. Second, the model often generates duplicate images without progress and cannot update the parameters efficiently. In this paper, we propose a two-stage model to solve the above problem. 1)We remove the redundancy structure and use an improved network structure that reduces the scale of the model size. 2)Our method employs a model trained in two stages instead of simultaneously, which shortens the training time and ensures that the model does not have vanishing gradients or mode collapse. In addition, we added an attention mechanism to the model to help optimize details. Experimental results show that our model saw excellent results in terms of generation quality and reduced model size on CUB(IS 4.83, FID 15.13) and COCO dataset(FID 33.74).