Generative Adversarial Network Technologies and Applications in Computer Vision

Table 6

Comparisons of GAN models in domains transfer.

GAN models	Improvements	Shortages	Applications
Pix2pix [71]	Using U-NET network and PatchGAN architecture which make the model easy to converge and images are realistic	The model is a supervised model and also needs the data with both tags and markeds	Style transfer and other applications
CycleGAN [72]	Cycle loss, self-constraint, and two-step transformation. The model training does not need a large data set	The quality of generated images is lower than pix2pix	Most of the style conversion scenes
DiscoGAN [73]	Using two GAN models to discover cross-domain relationships reducing model collapse and improve image quality	Data sets must be one-to-one paired images	Most scenes in domain transfer
StarGAN [74]	Adding control information of a domain to understand the image which domain does it belongs to	Needs a large number of different data sets	Multidomain transfer
DTN [75]	Using several complex loss functions, generating appealing emoji. From a facial image	The generated images with low quality	Using real photos to generate cartoon images