Generating Constrained Multi-target Scene Images Using Conditional SinGAN

As a variant of generative adversial network(GAN), SinGAN has been attracting a lot of research interest. Aiming at the disadvantage that SinGAN can easily generate multi-target images that do not conform to human logic, and the randomness of generating images is exsiting, simutaneously, the controllability of the generated images are poor, we have improved SinGAN’s architecture and loss function, and made it a conditional GAN instead of the original nonconditional GAN, so as to make the images more reasonable in spatial layout and semantic information. Conditional SinGAN realizes the function of manipulating the number and layout of the target in the image with the subjective will of the users. The experimental results show that conditional SinGAN has achieved ideal effect.