Prompt-Guided Aesthetic Assessment with Cross-Modal Semantic Alignment

The rapid proliferation of digital technology has led to an unprecedented surge in visual creations, making the automation of image aesthetic quality assessment increasingly imperative. However, challenges such as the limited scale of dataset samples, subjective annotation biases, and semantic ambiguity hinder progress in this domain. Moreover, traditional supervised learning paradigms rely excessively on the feature distributions of specific datasets, thereby compromising model generalization. To address these limitations, this paper proposes an image aesthetic quality assessment strategy based on prompt learning. By employing antonymous prompt text, annotation ambiguity is effectively mitigated, while an ensemble prompt reduces the bias introduced by individual prompts. Experimental results demonstrate that the proposed method achieves state-of-the-art EMD metrics on the AVA dataset and delivers exceptional performance across multiple publicly available datasets, significantly outperforming existing methods. Furthermore, cross-domain experiments conducted on diverse photography and fine art datasets underscore the model’s remarkable generalization capability.