An Acceleration Framework for High Resolution Image Synthesis

• DOI: 10.48550/arxiv.1909.03611
• Publication Date: 2019-01-01
• Authors: Jinlin Liu; Yuan Yao; Jianqiang Ren
• Keywords: Code (set theory)

Synthesis of high resolution images using Generative Adversarial Networks (GANs) is challenging, which usually requires numbers high-end graphic cards with large memory and long time training. In this paper, we propose a two-stage framework to accelerate the training process synthesizing images. High are first transformed small codes via trained encoder decoder networks. The code in latent space times smaller than original Then, train generation network learn distribution codes. way, generator only learns generate instead Finally, decode generated image networks so as output synthesized Experimental results show that proposed method accelerates significantly increases quality samples. acceleration makes it possible less limited hardware resource. After method, takes 3 days 1024 *1024 on Celeba-HQ dataset just one NVIDIA P100 card.