Abstract: Multimodal knowledge graphs serve as critical knowledge infrastructure for intelligent information systems. However, the continuous emergence of new entities in real-world applications poses critical challenges to existing knowledge retrieval-augmented reasoning systems, constraining the dynamic adaptability of intelligent information systems. Existing methods mainly focus on textual-visual semantic alignment, neglecting the constraining effect of graph topological structure on reasoning. Moreover, due to modality heterogeneity, multiple modalities are often difficult to collaboratively improve the reasoning performance on emerging entities. To this end, we propose SaCR, a Structure-Anchored Multimodal Contrastive Learning based Retrieval-Augmented Inductive Reasoning framework, which unifies multimodal representations with graph structure as anchors and enhances reasoning generalization. First, we propose entity neighbor anchor and relation anchor mechanisms to map graph structural information into a unified textual semantic space, achieving collaborative modeling of graph structure and textual-visual modalities. Second, through multimodal unified contrastive learning, we jointly optimize the structure anchors and multimodal encoders to shorten the distance between query and target entity representations. Finally, a retrieval-augmented inference mechanism is introduced to improve the generalization of reasoning on emerging entities. Experiments on 5 datasets under both inductive and transductive reasoning settings demonstrate that SaCR effectively improves the multimodal knowledge graph reasoning performance on emerging entities, providing complete and dynamic knowledge foundation for LLM-driven intelligent systems.