"Less is More" for AIGC Quality: A No-Reference IQA Model from Scratch
Abstract
Evaluating the perceptual quality of AI-generated content (AIGC) images without a reference is an open problem: generated images exhibit semantic inconsistency, hallucinated textures, and over-smoothing that differ fundamentally from the distortions targeted by classical no-reference IQA (NR-IQA). We present FineQA, a compact NR-IQA model trained from scratch on two public AIGC-IQA benchmarks that achieves CorrScore 0.7421 on a held-out 610-image test split, a +0.0595 gain over our initial 6.46 M-parameter generation (BaseIQA v1) while reducing parameters by 30.6× to just 211 K (0.84 MB). Three complementary contributions drive this result: (i) a BRISQUE-inspired MSCNQualityModule encoding multi-scale luminance statistics with only 208 parameters; (ii) self-supervised rotation-prediction pretraining at zero annotation cost; and (iii) 40-crop multi-scale test-time augmentation. FineQA's sub-megabyte footprint makes it directly deployable as a real-time quality filter in resource-constrained generative pipelines.