A Semantic Segment Encoder (SSE): Improving human face inversion quality through minimized learning space

Abstract

Recently, Generative Adversarial Networks (GAN) has been greatly developed and widely used in image synthesis. A Style-Based Generator Architecture for Generative Adversarial Networks (StyleGAN) which is the foremost, continues to develop human face inversion domain. StyleGAN uses insufficient vector space to express more than one million pixels. It is difficult to apply in real business due to distortion-edit tradeoff problem in latent space. To overcome this, we propose a novel semantic segment encoder (SSE) with improved face inversion quality by narrowing the size of restoration latent space. Encoder’s learning area is minimized to logical semantic-segment units that can be recognized by humans. The proposed encoder does not affect other segments because only one segment is edited at a time. To verify the face inversion quality, we compared with the latest encoders both Pixel2style2Pixel and RestyleEncoder. Experimental result shows that the proposed encoder improved distortion quality around 20% while maintain editing performance.