Editing Out-of-Domain GAN Inversion via Differential Activations

AbstractDespite the demonstrated editing capacity in latent space of a pretrained GAN model, inverting real-world images is stuck dilemma that reconstruction cannot be faithful to original input. The main reason for this distributions between training and data are misaligned, because that, it unstable inversion real image editing. In paper, we propose novel prior based framework tackle out-of-domain problem with composition-decomposition paradigm. particular, during phase composition, introduce differential activation module detecting semantic changes from global perspective, i.e., relative gap features edited unedited images. With aid generated Diff-CAM mask, coarse can intuitively composited by paired way, attribute-irrelevant regions survived almost whole, while quality such an intermediate result still limited unavoidable ghosting effect. Consequently, decomposition phase, further present deghosting network separating final fine reconstruction. Extensive experiments exhibit superiorities over state-of-the-art methods, terms qualitative quantitative evaluations. robustness flexibility our method also validated on both scenarios single attribute multi-attribute manipulations. Code available at https://github.com/HaoruiSong622/Editing-Out-of-Domain.