For decades, the field of Image Quality Assessment (IQA) has focused on a single, reductive goal: predicting a scalar score that correlates with human judgment. Whether it is a Mean Opinion Score (MOS) or a Fréchet Inception Distance (FID), these numbers tell us how much an image is degraded, but they fail to explain why. A score of 3.2 does not tell a photographer to fix the white balance, nor does it tell a medical professional if a CT scan is blurry due to patient motion or sensor noise.
A new research direction, often termed Explainable Image Quality Assessment (EIQA), is shifting the paradigm from black-box regression to linguistic reasoning. By leveraging Multi-modal Large Language Models (MLLMs), researchers are building systems that can diagnose specific distortions, localize artifacts, and provide actionable feedback in natural language.
The Failure of Scalar Scores
Traditional models like TReS, developed by S. Alireza Golestaneh and Kris Kitani, successfully combined CNNs for local features with Transformers for global context to achieve high correlation with human scores. However, as noted in the NTIRE 2023 Quality Assessment of Video Enhancement Challenge, these models lack diagnostic utility. They cannot quantify whether a video is poor due to over-sharpening or unnatural color saturation.
This lack of interpretability is particularly critical in specialized domains. In medical imaging, Jiyao Liu and colleagues argue in MedQ-Bench that scalar metrics lack the clinical context necessary for decision-making. Similarly, in speech processing, Chen Chen highlights that traditional audio metrics are "quality-blind," failing to distinguish between background noise and digital clipping in ALLD.
Bridging Pixels and Prose
The breakthrough in this field comes from treating quality assessment as a translation task: converting visual degradations into descriptive text. Zhiyuan You and Jinjin Gu pioneered this with DepictQA, which uses MLLMs to generate human-like linguistic descriptions of image quality. By training on hierarchical datasets like DQ-495K, their model, DepictQA-Wild, can reason about why one image is better than another, identifying specific texture damages or compression artifacts.
This transition works by aligning visual features with a "quality vocabulary." Instead of mapping an image to a point on a 1–5 scale, models like Q-Adapt, proposed by Xin Li and Haoning Wu, use progressive instruction tuning. This allows a model to first perceive universal distortions and then adapt to specific visual prompts, mimicking the human "Perception-Decision" process.
Disentangling Vibes from Pixels
A key insight in modeling aesthetic perception is that "quality" is not monolithic. Haoning Wu and Chaofeng Chen introduced DOVER, a framework that disentangles "vibes" (aesthetic appeal like composition and semantics) from "pixels" (technical quality like blur and noise). This separation is vital for User-Generated Content (UGC), where a video might have high aesthetic value despite technical flaws.
Similarly, in the realm of graphic design, the Agentic Design Review System by Sayan Nag uses a "design jury" of specialized AI agents. Each agent critiques a different dimension—such as color harmony or typography—to provide actionable feedback rather than a single score.
From Diagnosis to Actionable Guidance
The ultimate goal of EIQA is to move beyond diagnosis toward correction. Hanwei Zhu and Shiqi Wang proposed AgenticIQA, a modular framework where a "Planner" agent selects specialized tools to analyze an image, and a "Summarizer" provides a final interpretable report.
This reasoning-based approach enables "what-if" scenarios. For instance, Manxi Lin developed Diff-ICE, which uses diffusion models to generate counterfactual explanations for fetal ultrasound images. By showing a clinician a "standard plane" version of their low-quality scan, the model provides a visual roadmap for how to adjust the probe in real-time.
By replacing the "black box" with a "transparent critic," this research direction ensures that AI-driven quality assessment is not just accurate, but useful for the humans who rely on it.