Abstract: Infrared and visible image fusion aims to integrate thermal target information and visible texture details into a single image. Existing methods usually emphasize low-level texture preservation and contrast enhancement, while the consistency of high-level cross-modal semantics and the requirements of downstream tasks such as object detection are often less considered. In this paper, we propose SGMFusion, a semanticguided Mamba fusion network. The network uses dual ResNet-18 encoders to extract multi-scale features. For shallow features, a local spatial attention fusion module is adopted to preserve texture and edge details. For deep features, a Cross-Mamba Fusion module is designed to capture long-range cross-modal dependencies. The fused features are further processed by multi-scale semantic aggregation, and semantic cues are injected into the reconstruction path through a conditional-normalization-based semantic injection module. This design helps improve semantic consistency during image reconstruction. Meanwhile, a base-residual output strategy is introduced to improve fusion stability. During training, the fusion results are jointly constrained in terms of intensity, gradient information, structural similarity, and statistical correlation, thereby improving pixel fidelity, structural clarity, edge consistency, and statistical relevance. Comparative experiments against 7 state-of-theart methods on the MSRS, TNO, and LLVIP datasets show that SGMFusion achieves competitive performance on several fusion metrics. In the YOLOv8 object detection experiment, it also obtains the highest mAP among the compared methods, indicating that the proposed method is suitable for detection-oriented fusion tasks.