Adaptive Compression of Large Vision Models for Efficient Image Quality Assessment of AI-Generated Content

Shivam Bhardwaj, Tushar Shinde; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 3140-3147

The proliferation of Artificial Intelligence Generated Content (AIGC) demands scalable and efficient Image Quality Assessment (IQA) techniques, particularly for deployment in resource-constrained environments. Our approach leverages on the foundation models such as CLIP and DINO, but their computational overhead limits practical usage. In this work, we present a novel framework for efficient IQA by adaptively compressing Large Vision Models (LVMs) using layer-wise pruning and mixed-precision quantization guided by layer importance scores. Our method dynamically reduces model size while preserving perceptual sensitivity critical to assessing generative artifacts. We evaluate our approach on two dedicated AIGC IQA benchmark datasets: AGIQA-1K and AGIQA-3K and demonstrate up to 95% reduction in model size with minimal drop in human-perception correlation metrics such as SRCC and PLCC. Our method sets a new state-of-the-art in compact foundation models for generative content quality assessment and offers a scalable path toward real-time AIGC monitoring on edge platforms.