Image inpainting is a great challenge when reconstructed with realistic textures and required to enhance the consistency of semantic structures in large-scale missing regions. However, popular structural-prior guided methods rely mainly on the structural features, which directly accumulate and propagate random noise, causing inconsistencies in contextual semantics within the flled regions and poor network robustness.