Edit-discrepancy-guided feature transformation in encoder-based GAN inversion for real image attribute editing

Image attribute editing based on generative adversarial networks (GANs) typically begins by mapping real images to the latent space of a pretrained StyleGAN, followed by manipulating the corresponding latent codes. However, low-rate latent codes suffer from an information bottleneck, making it challenging to faithfully reconstruct complex real images. Recent encoder-based methods enhance reconstruction by injecting high-rate features into intermediate generator layers to better preserve fine details, but they often yield misaligned details in the edited images. The primary reason is that these methods still rely on global linear transformations of high-rate features, which overlook the nonlinear and spatially localized nature of real edits. To this end, we build on a high-fidelity encoder-based GAN inversion backbone and introduce an additional adaptive feature editor that is specifically trained to convert high-rate features during editing so that fine details are correctly aligned with the edited image. The backbone refines the feature through a cross-attention mechanism and residual enhancement. Building on this, the feature editor employs a window-based cross-attention mechanism to extract a discrepancy signal between the original and edited generator features, which specifies both where to modify and what content to change. This signal is then fused into the feature through spatially adaptive modulation techniques, enabling region-selective attribute changes while preserving irrelevant details. Experiments on face and car benchmarks demonstrate that our method improves both reconstruction fidelity and editing quality compared to existing GAN inversion methods.