DiffusionCom: Structure-Aware Multimodal Diffusion Model for Multimodal Knowledge Graph Completion

Multimodal knowledge graphs (MKGs) are widely used in recommendation systems, information retrieval, and question answering, but they remain highly incomplete, driving the need for multimodal knowledge graph completion (MKGC) methods. Most existing approaches rely on discriminative models that maximize conditional likelihood, neglecting the underlying data distribution and limiting their ability to capture complex real-world relations. To address this, we propose DiffusionCom, a structure-aware multimodal diffusion model for MKGC. DiffusionCom frames MKGC as generating the joint probability distribution of (head, relation) pairs and candidate tail entities from noise. We also propose Structure-MKGformer as its encoder, which captures structural information via a multi-modal graph attention network (MGAT) and adaptively fuses it with entity representations. Trained with both generative and discriminative losses, DiffusionCom outperforms state-of-the-art models on FB15k-237-IMG and WN18-IMG datasets—notably, achieving a 27.5% relative improvement in Hits@1 on FB15k-237-IMG.