Synthetic Code Mixing for StyleGAN Image Manipulation

Lately, there has been a proliferation of StyleGAN inversion methods aimed at facilitating facial editing. While many of these methods emphasize reconstruction quality and editability, the disentanglement of the inverted latent code has rarely been studied. In this work, we focus on both disentanglement and editability for facial editing. Our primary insight for achieving disentangled and editable StyleGAN inversion is to enforce a linear subspace for the inverted latent code. To this end, we introduce Synthetic Code Mixing (SCM), a method improves the disentanglement and editability of StyleGAN inversion from two aspects. Firstly, we leverage the synthetic latent code from the original <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathcal{W}$</tex> space to guide the encoder's output space towards <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathcal{W}$</tex>. This is built upon the fact that the original <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathcal{W}$</tex> space of StyleGAN is more linear. Secondly, we impose a linear change between two inverted latent codes, ensuring that the interpolated code aligns with the line connecting them. We provide comprehensive empirical evidence to demonstrate that our method achieves superior disentanglement and editability for StyleGAN inversion.