Existing dehazing methods often fail to jointly restore fine local textures and coherent global structure: haze obscures subtle details and breaks long-range spatial continuity. To address this, we propose a dual-branch architecture combining a ViT(Vision Transformers)-based global branch for capturing long-range dependencies with a CNN-based local branch for extracting detailed features. Central to our design is the Adaptive Global–Local Feature Fusion (AGLFusion) module, which employs content-aware feature selection and a gating mechanism to dynamically weight and integrate hierarchical cues across scales. By uniting the precision of CNNs in texture recovery with the contextual modeling of ViTs, our framework overcomes the classic trade-off between local-detail preservation and global-context restoration. Extensive evaluations on synthetic and real-world hazy datasets demonstrate that our method surpasses state-of-the-art dehazing techniques in both detail fidelity and structural coherence.