Multimodal Intent Recognition (MIR) aims to infer a speaker’s communicative intent by integrating heterogeneous textual, acoustic, and visual signals. However, intent-related cues are often distributed across fine-grained local variations and global semantic representations, making it challenging for existing approaches to explicitly model cross-modal and local-global dependencies within a unified framework. To address this challenge, we propose the Hybrid-scale Heterogeneous Graph Fusion Network (H$^2$GF-Net), which integrates affective semantic prior–guided local key information selection with progressive hybrid-scale heterogeneous graph modeling for multimodal intent understanding. Specifically, the model first performs global-scale multimodal affective semantic understanding by generating high-level affective-intent semantic descriptions, which serve as semantic priors beyond conventional emotion features. Guided by these priors, H$^2$GF-Net explicitly selects intent-relevant local units across textual, acoustic, and visual modalities, enabling precise localization of critical multimodal cues. Furthermore, a progressive hybrid-scale heterogeneous graph is constructed to capture dependencies between local and global representations as well as cross-modal interactions, and these graph representations are integrated through structure-aware graph fusion to support intent prediction. Experimental results demonstrate that the proposed method consistently outperforms state-of-the-art approaches on a benchmark dataset, while additional generalization evaluations further verify its robustness under extended evaluation settings.