Image inpainting aims to fill the missing regions naturally, and its essence lies in leveraging global context to reconstruct coherent structure and local features to restore fine textures. Existing Mamba-based techniques in computer vision primarily rely on corner-to-corner scanning strategies, which limit the interaction between an image’s central and peripheral regions. Besides, Mamba’s recursive processing of unfolded 1D image sequences disrupts images’ local structures. In this work, we propose InpaintingMamba, an image inpainting model that integrates Mamba’s global modeling capability with CNN’s strength in local feature extraction. Specifically, to ensure comprehensive information flow during inpainting, we design a periphery-center scanning strategy that establishes long-range dependencies from eight directions. Meanwhile, we employ parallel depth-wise convolutions to capture local details and textures. Extensive experimental results show that by combining the strengths of Mamba and CNN, InpaintingMamba achieves superior performance compared to SOTAs on both the CelebA-HQ and Places2 datasets. Project website is https://github.com/Lucas-image/InpaintingMamba