Embedding Chinese Face Painting Into the StyleGAN Latent Space

Pengsen Ma, Tao Xue

Abstract: We propose an efficient algorithm to embed China’s traditional Chinese painting into the latent space of StyleGAN for the first time in the presence of noise. This embedding allows us to complete the pluralistic image editing of Chinese painting style faces by using only the StyleGAN pretraining face model without the need to retrain new models. Compared to the previous reverse network StyleGAN-Encoder, the new training model can improve the image generation speed under noise by 10% and FID by approximately 21%. We for the first time propose the application of the deep residual shrinkage networks to the image generation problem and verify the effectiveness of the proposed method through experiments on various noises.