TKScore for generated Thangka image quality assessment with perceptual alignment and structural constraints

With rapid advances in generative models, generated image quality assessment has shifted from distribution-level evaluation toward instance-level and human-aligned assessment. However, existing metrics remain limited for generated Thangka images, whose quality depends on strict iconographic structures, symbolic color semantics, and fine decorative details. To address this challenge, we propose TKScore, a perception-aligned and structure-aware framework for generated Thangka image quality assessment. TKScore integrates adaptive region-weighted perceptual modeling for key iconographic regions, spatial-frequency collaborative feature extraction for texture and structural representation, and multi-dimensional attention regression for predicting color, texture, clarity, contour, and structure quality. Experiments on a self-built generated Thangka dataset show that TKScore outperforms representative traditional and modern baselines, achieving SRCC, PLCC, and KRCC values of 0.699, 0.728, and 0.516, respectively. These results demonstrate the effectiveness of TKScore for interpretable and human-aligned quality assessment of generated Thangka images.