Conditional Generation of Adversarial Networks Based on Multiple Generators

Conditional generative adversarial network are widely used in image translation and many other fields. However, traditional conditional generative adversarial networks have the problem of model collapse. To solve this problem, we proposed a conditional generative adversarial network model based on multiple generators. It uses multiple generators to obtain multiple outputs, and adds a distance constraint between multiple generators to output multimodal results. Experiments on Edges2Shoes and Facade datasets show that the diversity distance index LPILS between generated images can be effectively increased with our method. In addition, it also has good results in coloring application scenarios.