Facial Age Synthesis by StyleGAN2 with Attention

Tung-Ke Hsieh, Tsung-Jung Liu, Kuan-Hsien Liu

Abstract: The task of age synthesis involves capturing the changing appearance of a person over time. Efficiently representing such complex transformations on facial images is very challenging. We utilize the attention mechanism for enhanced feature extraction combined with expressive semantic latent space of StyleGAN2, which facilitates additional refinement and editing of the generated images, In addition, we incorporate an identity discriminator so that the identity of the generated images can be preserved. Finally, through qualitative and quantitative evaluations, we demonstrate that our method achieves better age synthesis accuracy compared to state-of-the-art methods.