Paper-Stylegan-Image2StylegGAN

Image2StyleGAN: How to Embed Images Into the StyleGAN Latent Space?

Accept: 2019ICCV

Problem：

基于优化(optimized-based)的StyleGAN反转

Framework

提出w+空间。即优化的latent code由原先stylegan的w变为w+(原先stylegan的w空间是通过复制然后送入styleblock进行调制，即18个w是完全相同的)

• LOSS
• perceptual loss and pixel-wise MSE loss
  $$w^{\ast }=\underset{w}{min}{L}_{p}ercept(G(w),I)+\frac{{\lambda }_{mse}}{N}{\Vert \begin{array}{c}G(w)-I\end{array}\Vert }_2^2$$
• perceptual loss
  $$L_{percept}(I_1,I_2)=\sum _{j=1}^4\frac{{\lambda }_j}{N_j}{\Vert \begin{array}{c}{F}_j(I_1)-F_j(I_2)\end{array}\Vert }_2^2$$
  perceptual loss的计算分别使用 VGG-16的$conv$1_1, $conv$1_2, $conv$3_2, $conv$4_2