Two-Stage Structure-Guided Residual Diffusion Model for Image Inpainting

In recent years, diffusion models have shown great potential in image inpainting but often suffer from structural inconsistencies and blurring. To address this, we propose SGRDM, a two-stage approach integrating a structure-guided residuals diffusion model. The first stage restores structural components using an edge-preserved smoothed image, while the second stage leverages the restored structure to guide texture inpainting through residual learning in a diffusion model. This method enhances texture quality, accelerates convergence, and improves training efficiency. Experiments on CelebA-HQ and Paris Street View datasets demonstrate superior structural integrity and outstanding performance in both qualitative and quantitative evaluations.