ThemeIQA:Theme-Oriented Generated Image Quality Assessment

With the rapid advancement of generative AI, large-scale text-to-image models are widely used in content creation. Creative demands are shifting from single images to serialized expressions, leading to a rise in groups of consecutively generated images with a shared semantic theme. However, assessing theme-oriented quality within such image groups remains challenging. Current AIGC quality assessment methods are predominantly tailored for individual images or consecutive video frames, rendering them inadequate for evaluating image groups. To address this challenge, this paper proposes ThemeIQA, a no-reference AIGC image quality assessment method tailored for theme-conditioned image set, with the goal of predicting overall theme-oriented quality. ThemeIQA jointly considers both single-image and cross-image perspectives, and provides fine-grained scores along four dimensions: perceptual quality, text-image alignment, subject consistency, and scene richness. Specifically, the subject consistency branch integrates text-conditioned subject features with intra-group similarity features to construct a group-level similarity-augmented representation for prediction. In addition, to mitigate the negative impact of repetitive scene content on visual perception, the scene richness branch introduces a discrepancy measurement module to characterize the extent of background variation among images within the group. Furthermore, to validate the proposed method, a large-scale theme-conditioned image quality assessment database, TAIGCIQA-10K, is constructed. The database is generated using eight text-to-image models based on single-subject and multi-subject prompt groups, where the subject categories span diverse types. Experimental results demonstrate that, at both the overall and fine-grained metrics, ThemeIQA yields assessments that exhibit stronger alignment with subjective ratings and demonstrates superior performance relative to existing methods.