FineQA: A Lightweight No-Reference Quality Assessment Model for AI-Generated Images under Severe Compute and Data Constraints

Most no-reference image quality assessment (NR-IQA) methods for AI-generated content (AIGC) depend on large, externally pretrained backbones, resources frequently unavailable under the compute, licensing, and data constraints common to institutions in the Global South. We present FineQA, a fully-from-scratch NR-IQA model with 210,537 parameters (0.84 MB) trained on only approximately 4,060 annotated images with no external pretraining, on a single entry-level GPU in under an hour. On a held-out 610-image AIGC-IQA test split, FineQA reaches CorrScore = 0.7421 (40-crop test-time augmentation, TTA), improving on a from-scratch reference baseline (0.5875) by +0.1546 while using 19x fewer parameters. Three design choices drive this: (i) a BRISQUE-inspired MSCN Quality Module adding only 312 parameters; (ii) self-supervised pretraining that eliminates early training collapse; and (iii) multi-scale TTA whose gain saturates at 20 crops (CorrScore = 0.7445). A pretext-task study further shows jigsaw spatial-layout pretraining lifts CorrScore to 0.7519, outperforming rotation, distortion-type prediction, and random initialization, despite the lowest pretext-task accuracy of the three supervised alternatives. We argue this budget-first design philosophy, rather than scaling up pretrained backbones, is a more accessible and reproducible path to AIGC quality assessment in low-resource research settings.