VGIQA: Fine-Grained Image Quality Assessment with Visual Grounding Rewards

Weichang Wang; Liyu Chen; Shuyue Guan; Jie Hu; Fei Gao; Wei Li; Nannan Wang

cs.CVOct 27, 2026

VGIQA: Fine-Grained Image Quality Assessment with Visual Grounding Rewards

Authors: Weichang Wang, Liyu Chen, Shuyue Guan, Jie Hu, Fei Gao, Wei Li, Nannan Wang

Organizations: State Key Laboratory of Integrated Services Networks, Xidian University · Huawei Foundation Model Department

Abstract

Image quality assessment (IQA) aims to evaluate the perceptual visual quality of images. Recently, the rapid development of Multi-modal Large Language Models (MLLMs) has significantly advanced the IQA field. By integrating chain-of-thought (CoT) reasoning and reinforcement learning (RL), these models can generate more accurate and interpretable results while relying on less training data. However, existing MLLM-based methods are predominantly constrained to holistic, image-level evaluations and lack the capability to produce fine-grained, localized assessments, which substantially limits their practical applicability. To address this limitation, we propose VGIQA, a novel IQA model endowed with both local and global quality assessment capabilities. VGIQA is trained via a two-stage paradigm comprising a cold-start phase followed by reinforcement learning. Driven by carefully designed local quality assessment reward functions, our model exhibits exceptional image quality understanding on VGSR-2K, a new dataset we propose specifically for local quality evaluation, significantly outperforming existing open-source MLLMs such as Qwen2.5-VL-72B-Instruct and Qwen3-VL-32B-Instruct. Furthermore, the local quality evaluation task can effectively enhance the model's performance on the degradation perception task.

Figures & tables

Figure 1: Qualitative visualization of the local quality assessment results. Given a reference image, a distorted image to be evaluated, and a user query, VGIQA produces more reliable and accurate responses compared with the state-of-the-art RL-based IQA method Q-insight and the general large vision-language model Qwen2.5-VL-72B-Instruct [ 1 ] .

Figure 2: Example applications of the proposed VGIQA framework. Built upon a Qwen-based foundation model, VGIQA achieves comprehensive quality understanding via multi-task reinforcement learning. While existing models (e.g., Q-Insight) are limited to holistic Score Regression and Degradation Perception , VGIQA explicitly bridges the gap in fine-grained Local Quality Assessment . Driven by novel visual grounding rewards, it adopts a rigorous "look and then judge" paradigm—precisely localizing spatial artifacts before formulating final judgments.

Figure 3: Overview of the proposed VGIQA framework. The framework follows a two-stage training paradigm. Stage 1 (Cold Start): The Qwen-based foundation model is initialized via SFT using CoT reasoning data distilled from Qwen3-VL-Plus [ 2 ] to establish structural formatting and inject IQA priors. Stage 2 (Multi-Task GRPO): Reinforcement learning jointly optimizes Local Quality Assessment, Degradation Perception, and Score Regression. Alongside a shared format reward, task-specific rewards are dynamically routed: a dedicated local reward exclusively supervises the local task (computing zero for non-local tasks), evaluating spatial bounding boxes via Hungarian Matching, negative DIoU, and F1 Score. Meanwhile, global tasks are optimized via verifiable score and classification/intensity rewards, unifying local-and-global quality evaluation.

Figure 4: Overview of the VGSR-2K dataset construction pipeline. The dataset is curated via a rigorous human-in-the-loop pipeline. (1) Candidate Generation: Qwen3-VL-Plus [ 2 ] and Seed1.6-VL independently process LR-SR image pairs, yielding six candidate spatial bounding boxes. (2) Double-Blind Selection: Two annotator groups select the most consistent candidate; consensus establishes the baseline, while divergence prompts expert resolution. (3) Double-Blind Correction & Merging: Annotators refine baselines by adding or removing boxes. Corrections are merged via the Hungarian algorithm with a negative DIoU cost metric, extracting the MBR for matched pairs. (4) Final Review & Balancing: After expert review, pristine LR-HR image pairs are introduced as negative samples (with empty annotations) to calibrate spatial predictions and prevent trivial over-generation.

Table 5

Method	Metrics	Blur	JPEG	Darken	Noise	Null	Average
Q-Insight [ 18 ]	Deg.Acc.	91.48%	98.02%	63.23%	16.13%	64.93%	66.57%
	Lev.Acc.	43.05%	46.04%	22.70%	6.91%	-	34.23%
VGIQA (Ours)	Deg.Acc.	94.17%	98.51%	67.71%	62.21%	94.03%	82.18%
	Lev.Acc.	52.91%	44.55%	23.77%	23.96%	-	43.94%

Table 3: Quantitative comparison of the Degradation Perception task. VGIQA significantly improves the perception accuracy of global physical degradations, especially in Noise and Null(no degradation). The best and second-best results are highlighted in bold and underlined , respectively.

Method	KonIQ	SPAQ	KADID	PIPAL	LIVE-Wild	AGIQA-3K	CSIQ	Average
NIQE [ 13 ]	0.533/0.530	0.679/0.664	0.468/0.405	0.195/0.161	0.493/0.449	0.560/0.533	0.718/0.628	0.521/0.481
BRISQUE [ 22 ]	0.225/0.226	0.490/0.406	0.429/0.356	0.267/0.232	0.361/0.313	0.541/0.497	0.740/0.556	0.436/0.369
NIMA [ 36 ]	0.896/0.859	0.838/0.856	0.532/0.535	0.390/0.399	0.814/0.771	0.715/0.654	0.695/0.649	0.697/0.675
MUSIQ [ 15 ]	0.924/0.929	0.868/0.863	0.575/0.556	0.431/0.431	0.789/0.830	0.722/0.630	0.771/0.710	0.726/0.707
CLIP-IQA+ [ 37 ]	0.909/0.895	0.866/0.864	0.653/0.654	0.427/0.419	0.832/0.805	0.736/0.685	0.772/0.719	0.742/0.720
ManIQA [ 45 ]	0.849/0.834	0.768/0.758	0.499/0.465	0.457/0.452	0.849/0.832	0.723/0.636	0.623/0.627	0.681/0.658

Table 4: Quantitative comparison of the score regression task between our VGIQA and other competitive IQA methods. Results are reported in the format of PLCC/SRCC. All methods except handcrafted ones are trained on the KonIQ dataset. The best and second-best results are highlighted in bold and underlined , respectively.

Method	Metrics	Blur	JPEG	Darken	Noise	null	Average
Global-only	Deg.Acc.	51.12%	97.52%	43.05%	43.78%	72.39%	59.96%
	Lev.Acc.	25.11%	45.05%	14.08%	27.19%	-	33.63%
Global + Local (Ours)	Deg.Acc.	94.17%	98.51%	67.71%	62.21%	94.03%	82.18%
	Lev.Acc.	52.91%	44.55%	23.77%	23.96%	-	43.94%

Table A : Ablation study of multi-task training on the degradation perception task. The Global tasks include score regression and degradation perception. Joint training with the local quality assessment task significantly improves global degradation perception. The best and second-best results are highlighted in bold and underlined , respectively.

Figure A : Qualitative illustration of the candidate generation process. This figure presents a given $[I_{Reference},I_{Generated}]$ image pair alongside its corresponding six different candidate generation results. By providing such a diverse set of candidate bounding boxes, human annotators are empowered to conduct more comprehensive screening and rigorous cross-correction in the subsequent annotation stages.

System Prompt (Shared)

A conversation between User and Assistant. The user asks a question, and the Assistant solves it. The assistant first thinks about the reasoning process in the mind and then provides the user with the answer. The reasoning process and answer are enclosed within <think> </think> and <answer> </answer> tags, respectively, i.e., <think> reasoning process here </think><answer> answer here </answer> .

Prompt for Local Quality Assessment (Ours)

You are an expert in super-resolution (SR) image quality assessment. You are provided with:

First image: Low-Resolution (LR) reference image (ground truth basis)

Second image: Super-Resolved (SR) output image

Your task is to carefully compare the two images to detect, localize, and evaluate all regions in the SR image where the generation is problematic (e.g., hallucinations, distortions, or objective degradations).

Table B : Complete prompt templates used in our multi-task joint training framework. The system prompt, degradation perception prompt, and score regression prompt are directly adopted from Q-Insight [ 18 ] , while the local quality assessment prompt is our newly designed core contribution.

Image Quality Assessment

Image Quality Assessment measures perceived image quality, distortion, and usefulness with computational metrics and human-aligned evaluation.

See all 175 papers

From Glance to Scrutiny: Progressive Distortion Reasoning for Fine-Grained Image Quality Assessment
Sep 8, 2026Aoting Zhang, Mingze Gao, Dongbao Yang +5Image Quality Assessment
IIE, Chinese Academy of Sciences · University of Chinese Academy of Sciences · Xiaohongshu Inc. +1
Multi-Granularity Reasoning for Image Quality Assessment via Attribute-Aware Reinforcement Learning to Rank
Apr 7, 2026Xiangyong Chen, Xiaochuan Lin, Haoran Liu +3Image Quality Assessment VLM Reasoning
Henan Polytechnic University
Tool-IQA: Augmenting Image Quality Assessment with Simple Tools
Jun 15, 2026Guanyi Qin, Junjie Zhang, Chunming He +4Tool Use in VLMs Image Quality Assessment
1National University of Singapore · 2OPPO Research Institute · 3Nanyang Technical University +3

Sep 8, 2026J/K move · Enter open · S save

From Glance to Scrutiny: Progressive Distortion Reasoning for Fine-Grained Image Quality Assessment

Aoting Zhang, Mingze Gao, Dongbao Yang, Longyi Chen+4

IIE, Chinese Academy of Sciences · University of Chinese Academy of Sciences · Xiaohongshu Inc.+1

Image Quality Assessment

Open paper AskPDF arXiv

Multi-modal large language models (MLLMs) have demonstrated significant potential in image quality assessment (IQA) by bridging visual perception with descriptive evaluations. However, existing approaches mainly focus on holistic quality prediction, often functioning as black boxes that provide limited insight into where distortions occur and how they affect perceived quality, hindering fine-grained analysis of localized and heterogeneous degradations. We propose GS-IQA, a framework that reformulates IQA as a progressive Where--What--How diagnosis, emulating the human perceptual process from an initial glance to closer scrutiny. Since a severity judgment is meaningful only for a correctly localized and recognized region, we realize this progression through a two-stage reinforcement learning paradigm that respects such dependencies: the glance stage uses a perception-gated reward to establish where degradations lie and what they are, activating severity feedback only once both are correct, while the scrutiny stage introduces online reward-conditioned degradation generation to synthesize hard examples targeted at the model's perceptual bottlenecks, sharpening its discrimination of subtle severity variations. To enable systematic evaluation, we construct Diag-Bench, a region-level IQA benchmark of about 25K curated samples spanning 12 distortion types and five ordinal severity levels. Extensive experiments show that GS-IQA consistently surpasses state-of-the-art methods in distortion localization, recognition, and severity estimation, and that its diagnostic representations transfer effectively to conventional global quality prediction across diverse external benchmarks. Code and data will be released.

Abstract

Image Quality Assessment

Explore similar work

From Glance to Scrutiny: Progressive Distortion Reasoning for Fine-Grained Image Quality Assessment

Multi-Granularity Reasoning for Image Quality Assessment via Attribute-Aware Reinforcement Learning to Rank

Tool-IQA: Augmenting Image Quality Assessment with Simple Tools