FineQA: A Lightweight No-Reference Quality Assessment Model for AI-Generated Images via Self-Supervised Pretraining

We present FineQA, a fully-from-scratch no-reference image quality assessment (NR-IQA) model with 210,537 parameters (0.84MB) that reaches CorrScore=0.7421 (40-crop test-time augmentation, TTA) and 0.7257 (single crop) on a held-out 610-image AIGC-IQA test split, improving on a from-scratch reference baseline (0.5875) by +0.1546 while using 19× fewer parameters. A post-hoc TTA sweep on the identical model identifies 20 crops as the practical operating point (CorrScore =0.7445). Three design choices drive these results: (i) a BRISQUE-inspired MSCNQualityModule extracting multi-scale luminance statistics with only 312 extra parameters; (ii) self-supervised pretraining that eliminates early training collapse; and (iii) multi-scale TTA whose gain saturates at 20 crops. A six-variant pooling-head ablation and a four-condition pretext-task study provide isolated analyses; the latter shows that jigsaw spatial-layout pretraining further lifts CorrScore to 0.7519, outperforming rotation, distortion-type prediction, and random initialization– despite having the lowest pretext-task accuracy of the three supervised alternatives