Architectural Image Editing Based on Residual Spatial Networks

The complexity of the hierarchical structure of architectural images brings certain challenges to architectural image editing. Some works combine reconstruction and editing for architectural image editing, however, low-dimensional encoding in hidden space often results in loss of spatial information, causing generated images to be distorted and making it challenging to maintain structural consistency. To address these issues, we propose an architectural image editing model based on StyleGAN, named ABEditor. Specifically, we introduce a multi-level residual spatial network to learn multi-layer spatial and style features, which are connected to the corresponding dimensions in StyleGAN’s high-dimensional feature space. The input image can be accurately reconstructed through layer-wise correction while preserving rich spatial information. An encoder in the early layers of the synthesis network maps images to low-dimensional latent space, ensuring fidelity in color details. Additionally, to more accurately evaluate editing performance, we propose the average pixel edit distance based on segmentation (APEDS) metric. Comparative evaluations on the LSUN Church dataset show that the ABEditor framework improves in retaining building image details and spatial structure, outperforming related methods by 28.64% in SSIM, 20.74% in FID, 13.67% in L₂, 1.67% in LPIPS, 10.01% in PSNR, and 21.26% in APEDS.