<p>Single Domain Generalized Object Detection (SDGOD) is a challenging task that frequently arises in real-world applications. Despite recent advances in Vision Foundation Models (VFMs), existing methods often overlook semantic biases at the instance level or inadequately handle multimodal feature fusion. To address these limitations, we propose SemClip, a CLIP-based SDGOD model that leverages style enhancement and semantic consistency. Specifically, we design a dynamic fine-grained style enhancement strategy to generate diverse style information, enhancing the model’s generalization capability. By efficiently fusing visual and textual features using CLIP, we increase feature diversity. Additionally, we introduce a semantic consistency loss to minimize cross-domain semantic feature distribution bias and suppress interference from confusing attributes at the instance level. Experiments on multiple target domains validate the effectiveness and superiority of our proposed method. Here, we show that SemClip significantly outperforms existing SDGOD methods, achieving optimal performance on challenging domains such as Daytime Foggy and Dusk Rainy. Our work contributes to the broader field of domain generalization by providing a robust framework for cross-domain object detection. The source codes can be found at <a href="https://github.com/Yurunyi11111/SemClip.">https://github.com/Yurunyi11111/SemClip.</a></p>