Unveiling Visual Features in Artwork Classification: Towards Explainable Vision Transformers in the Arts

Recent advances in deep learning have enabled accurate artwork classification using models such as Vision Transformers (ViTs). However, interpreting the internal mechanisms behind such decisions remains challenging, especially in the abstract and symbolic domain of visual arts. We propose an interpretability framework that combines feature visualization via activation maximization with natural language grounding through a Multimodal Large Language Model. Our method extracts class-specific visual patterns learned by ViTs, synthesizes prototype images that activate key features, and generates human-readable descriptions. Applied to a large-scale art dataset, the approach reveals that ViTs attend to subtle and abstract cues—such as texture, shape, and composition—differently from natural image tasks. The resulting visual and textual explanations offer valuable insight into model behavior and move toward more transparent, human-aligned AI systems for art analysis.