Aesthetic evaluation of artistic images remains a challenging prediction problem, constrained by the subjective nature of human aesthetic cognition and the complex interplay between low-level visual structures and high-level semantic characteristics. In terms of model construction, the major bottleneck is to establish discriminative representations that can sufficiently model nonlinear dependencies across diverse heterogeneous modalities. In this paper, we propose VPT-IAA, a unified vision-language framework that formulates artistic aesthetic assessment as a cross-modal representation learning problem with explicit interaction modeling. The visual modality is encoded through a permutation-based feature transformation that enables structured aggregation of spatial information across multiple dimensions, while structured multi-dimensional aesthetic attribute descriptions are automatically generated by a MiniCPM-V multimodal large language model pipeline that decomposes artistic appreciation into seven independent dimensions across four hierarchical perception layers, and these descriptions are subsequently embedded using a task-oriented Transformer encoder to obtain attribute-consistent textual representations. To model cross-modal dependencies, we introduce a three-pathway attention mechanism that establishes complementary query-key-value interactions between visual and textual features, yielding a coupled representation space. In addition, bilinear pooling is employed to characterize second-order correlations between modalities, allowing the model to capture higher-order aesthetic relationships. Model hyperparameters are optimized via differential evolution to enhance stability and robustness. Experimental evaluation on the LAPIS dataset demonstrates that the proposed formulation achieves accurate and consistent aesthetic prediction, attaining an MAE of 1.796 and PC of 0.9827-representing a 15.8% reduction in MAE relative to the second-best method AesExpert and a 68.6% reduction relative to CNN-based baselines. Further analyses indicate that explicit cross-modal interaction modeling plays a dominant role in performance improvement, and the proposed framework maintains stable behavior across a wide range of artistic styles, including both representational and abstract artworks.
Zhang et al. (Thu,) studied this question.