Semantic-Aware GAN Manipulation for Human Head Editing

In recent years, there has been a rapid growth in the quality of image generation. It is mainly due to the development of generative models based on approaches from deep learning. At the moment, generative adversarial networks show themselves best. They can be used to synthesize data that is very similar to real data. However, the generation process itself often occurs without any control. Accordingly, one of the open and important questions is how to make the synthesis process controllable. Thus, it will be possible to semantically edit the output of the generative model. This will allow you to change, for example, shape, attributes, color of objects in the resulting image. Research on this topic in computer vision suggests that the result of generative models strongly depends on the hidden subspace from which the image is generated. This paper is devoted to a parsing of approaches that consider ways of controlling the inner space of generative-adversarial networks. Emphasis is placed on the consideration of the human head editing problem. The relevant data collections required for training generative-adversarial networks are described. Also, experiments were conducted on a part of the considered approaches, and conclusions on the results of the work are made.