Enhancing Image Inpainting via Multi-Resolution Fusion and Semantic-Aware Diffusion Models

Image inpainting is a critical technique in computer vision and image processing, aiming to restore missing or occluded regions in images. Despite advances, existing methods often struggle with coordinating global and local information, leading to semantic inconsistencies and abrupt transitions between repaired and surrounding regions. To address these challenges, we propose a novel multi-resolution image inpainting method based on a denoising diffusion probabilistic model (DDPM). Our approach integrates low-resolution image inpainting, which captures global structural information, with high-resolution inpainting that emphasizes fine-grained local details. By fusing these results, we achieve high-quality, semantically consistent inpainted images. Additionally, we introduce a Prompt-Aware Contextual Attention (PaCA) module to enhance the consistency between complementary content and textual cues, and a Frequency-Aware Diffusion Module (FADM) based on Fast Fourier convolution to capture high-frequency details in low-resolution images. Experimental results demonstrate superior performance in both qualitative and quantitative evaluations, achieving an FID score of 7.14 and an LPIPS score of 0.043 on the CelebA-HQ dataset, outperforming state-of-the-art methods. The semantics-aware diffusion model proposed in this paper, by introducing a multi-resolution feature interaction mechanism, effectively addresses the problem of semantic drift in complex scenarios that plagues traditional methods.