Enhancing Artwork Style Clustering via Neural Representation Re-Alignment

Abhishek Dangeti, Pavan Gajula, Vikram Jamwal, Vivek Srivastava; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 446-455

Style-based clustering facilitates the analysis of artistic style evolution, enables discovery of fine-grained stylistic variations, and reveals hidden connections within art collections. However, current neural style representations fail to adequately capture the inherently subjective nature of artistic style, resulting in suboptimal clustering performance and limited applicability to diverse style definitions. We introduce methods that leverage human preference feedback to enhance style-based clustering through representation alignment. Through comprehensive evaluation, we demonstrate that preference-guided alignment substantially improves clustering quality, while revealing that effectiveness is critically dependent on both the underlying representation architecture and the specific style definition employed. We contribute empirical insights into these architectural and definitional dependencies, and establish practical guidelines for selecting optimal representations across different clustering scenarios.