MGDT Enhances Multimodal Knowledge Graph Completion with Diffusion Transformers

Multimodal Knowledge Graph Completion (MKGC) involves inferring missing entities by integrating structural, textual, and visual information. Traditional diffusion-based MKGC methods often struggle because their denoisers must simultaneously handle relation-dependent cue selection, cross-modal semantic alignment, and structure-aware entity generation, leading to noisy and inconsistent conditions. To overcome these challenges, researchers propose MGDT: MLLM-Guided Diffusion Transformer with Relation-Adaptive Mixture-of-Experts. MGDT employs an "align-then-diffuse" paradigm. First, a Relation-Adaptive Semantic Routing Mixture-of-Experts (RASR-MoE) module intelligently selects relevant multimodal semantic transformation paths, effectively suppressing irrelevant modality interference. Second, a frozen Multimodal Large Language Model (MLLM) acts as a semantic anchor, aligning these routed multimodal representations into a unified latent space, thereby reducing cross-modal semantic heterogeneity. Finally, a Knowledge Graph Diffusion Transformer (KGDT) performs graph-conditioned denoising generation within this aligned space to produce the representation of the missing entity. Experimental results across three benchmark datasets consistently show that MGDT significantly outperforms existing strong baselines, demonstrating its effectiveness in MKGC.