Multimodal knowledge graph completion (MMKGC) aims to extract knowledge from real world data by structuring triple relations and integrating multimodal information. However, modality differences often lead to fragmented, misaligned representations, making integration and inference challenging. Existing deep learning-based fusion methods fail to preserve modality specific features and require significant computational resources. While tensor decomposition approaches help reduce resource use, they often suffer from overfitting, limiting their effectiveness. To address these issues, we propose MGFHD, a framework using hypersphere geometry, and geodesic interpolation to balance multimodal representations while extending the DUality-induced RegulArizer (DURA) to mitigate overfitting. Our experiments on four MMKG datasets demonstrate that MGFHD outperforms 19 baselines in both efficiency and predictive accuracy.