Visual-TransGNN is proposed as a Transformer-based Graph Neural Network for personalized fashion recommendation. The model enhances traditional multi-head attention with similarity-aware weighting and edge-contextual refinement to better capture structural and visual relationships between fashion items. Clothing images are embedded with pre-trained CNNs and structured into graphs based on visual and semantic similarity. The resulting latent representations integrate both aesthetic compatibility and user preferences, resulting in more accurate and diverse outfit recommendations. Experimental results on two public datasets–Vibrent and H&M–demonstrate improved performance over existing GNN-based models while offering enhanced interpretability through refined attention mechanisms. This work underscores the value of visual-aware attention in graph-based recommendation systems.