Descriptive Video Degradation Assessment with Multimodal Large Language Models

Abstract

We investigate descriptive video degradation assessment (VDA) for granular diagnosis beyond the interpretability limits of conventional video quality assessment (VQA). Unlike scalar mean opinion score (MOS) regression, descriptive VDA identifies spatio-temporal degradations and explains their collective perceptual impact, turning black-box judgments into actionable feedback for video enhancement, compression optimization, and AIGC refinement. Adapting multimodal large language models (MLLMs) to this task is hindered by scarce quality-specific video-text supervision and the tension between fine-grained spatial perception and long-range temporal reasoning. We address both challenges with score-anchored semantic synthesis, which uses MOS as semantic anchors for consistent quality-oriented descriptions, and hierarchical spatio-temporal adaptation, which aligns spatial degradation primitives on images before learning temporal quality dynamics from videos. Experiments show more accurate spatial attributions and more stable temporal judgments than prior VQA baselines, demonstrating strong diagnostic utility. A preliminary extension study further suggests that the framework can generalize beyond technical degradations to aesthetic, temporal, and AIGC-specific quality dimensions. Code and data will appear at Descriptive-VDA.