Comparison of Controllable Image Generation Methods for Face Synthesis

Generative Adversarial Networks (GAN) is widely used in the field of image generation because they can synthesize images reflecting various properties such as color, edges, drawing style, or background. In particular, GAN excels at realistically synthesizing faces, and they have had great success in manually controlling face attributes. However, when features extracted from face images are entangled, failure cases still occur during image generation. In this paper, we select two representative methods that can successfully solve these problems. We then analyze their strengths and weaknesses by direct performance comparison on CelebA. In these experiments, we identified which parts of the model are key to controlling face attributes when generating images.