DeepGeoFusion: Personalized Facial Beauty Prediction Through Geometric-Visual Fusion

Kunwei Wang^1*Dong Huang^2*Junmei Feng³Xiaoyi Feng¹

• ¹Northwestern Polytechnical University, Xi'an, China
• ²Air Force Medical University, Xi'an, China
• ³Xidian University Guangzhou Institute of Technology, Guangzhou, China

Personalized facial beauty prediction has emerged as a critical advancement beyond population-level models, with transformative applications in aesthetic surgery planning and user-centric recommendation systems. However, contemporary methods face three fundamental limitations: (1) inadequate modeling of aesthetic-sensitive facial regions, (2) ineffective fusion of heterogeneous geometric and visual features, and (3) high dependency on extensive annotation for personalization. To address these challenges, we introduce DeepGeoFusion – a novel framework that synergizes global deep visual features (extracted via Vision Mamba) with anatomically constrained facial graphs constructed from 86 landmarks through Delaunay triangulation. Our core innovation, the Graph Node Attention Projection Fusion (GNAPF) block, dynamically aligns cross-modal representations using topology-aware attention mechanisms. Crucially, DeepGeoFusion incorporates a lightweight adaptation mechanism that generates personalized preference vectors from only 10 seed images via confidence-gated optimization. Extensive experiments on SCUT-FBP5500 demonstrate that our geometric-visual fusion paradigm achieves statistically significant improvements in personalized prediction accuracy, while maintaining robustness across genders and ethnicities compared to state-of-the-art methods.