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

Text-to-image generative models produce failure modes, hallucinated textures, locally incoherent structures, semantic implausibility, that pixel-level metrics and natural-image no-reference image quality assessment (NR-IQA) methods are not designed to capture. Existing approaches to AI-generated-content IQA (AIGC-IQA) typically rely on large pretrained backbones, which are unavailable under tight capacity and data budgets. We study the under-explored from-scratch regime, training an NR-IQA model with no pretrained weights under a strict low-data constraint (~4K annotated images), where model capacity must be carefully budgeted and naive training is prone to collapse.