The paper "Structure Matters: Tackling the Semantic Discrepancy in Diffusion Models for Image Inpainting" addresses the issue of semantic discrepancy between masked and unmasked regions in image inpainting using denoising diffusion probabilistic models (DDPMs). The authors propose a novel structure-guided diffusion model named StrDiffusion to improve the consistency and meaningfulness of the inpainted results. The key contributions of the paper are: 1. **Structure-Guided Denoising Process**: StrDiffusion reformulates the conventional texture denoising process under the guidance of the structure, derived from the unmasked regions. This approach aims to leverage the time-dependent sparsity of the structure to guide the texture denoising process, ensuring consistent semantics in the early stages and reasonable semantics in the late stages. 2. **Adaptive Resampling Strategy**: An adaptive resampling strategy is introduced to monitor and regulate the semantic correlation between the denoised texture and the structure. This strategy dynamically adjusts the guidance from the structure to the texture based on the semantic correlation, enhancing the overall performance. 3. **Experiments and Results**: Extensive experiments on various datasets (Paris StreetView, CelebA, Places2) demonstrate the effectiveness of StrDiffusion. The results show that StrDiffusion significantly reduces the semantic discrepancy and improves the quality of the inpainted results compared to state-of-the-art methods. The paper provides a comprehensive analysis and validation of the proposed approach, highlighting its advantages over existing methods in handling semantic discrepancies and achieving more consistent and meaningful inpainted images.The paper "Structure Matters: Tackling the Semantic Discrepancy in Diffusion Models for Image Inpainting" addresses the issue of semantic discrepancy between masked and unmasked regions in image inpainting using denoising diffusion probabilistic models (DDPMs). The authors propose a novel structure-guided diffusion model named StrDiffusion to improve the consistency and meaningfulness of the inpainted results. The key contributions of the paper are: 1. **Structure-Guided Denoising Process**: StrDiffusion reformulates the conventional texture denoising process under the guidance of the structure, derived from the unmasked regions. This approach aims to leverage the time-dependent sparsity of the structure to guide the texture denoising process, ensuring consistent semantics in the early stages and reasonable semantics in the late stages. 2. **Adaptive Resampling Strategy**: An adaptive resampling strategy is introduced to monitor and regulate the semantic correlation between the denoised texture and the structure. This strategy dynamically adjusts the guidance from the structure to the texture based on the semantic correlation, enhancing the overall performance. 3. **Experiments and Results**: Extensive experiments on various datasets (Paris StreetView, CelebA, Places2) demonstrate the effectiveness of StrDiffusion. The results show that StrDiffusion significantly reduces the semantic discrepancy and improves the quality of the inpainted results compared to state-of-the-art methods. The paper provides a comprehensive analysis and validation of the proposed approach, highlighting its advantages over existing methods in handling semantic discrepancies and achieving more consistent and meaningful inpainted images.