MetaRefine: Score Refinement with Multi-Level Features for Image Assessment

Chuanzhi Xu; Chengkun Yue; Haolan Guo; Xuanhua Yin; Boyu Wei; Zihan Deng; Qiang Qu; Weidong Cai

cs.CVOct 29, 2026

MetaRefine: Score Refinement with Multi-Level Features for Image Assessment

Authors: Chuanzhi Xu, Chengkun Yue, Haolan Guo, Xuanhua Yin, Boyu Wei, Zihan Deng, Qiang Qu, Weidong Cai

Organizations: 1The University of Sydney · 2Indiana University · 3The University of Hong Kong

Abstract

An image's original dimensions, aspect ratio, and composition provide important cues that can affect its technical and aesthetic quality. Image assessment methods use resizing, cropping, padding, or patch sampling to accommodate different input sizes. However, these operations can alter or obscure original geometric and compositional cues. Existing assessment models also encode rich multi-level visual evidence that can be further extracted and exploited to improve existing models. This motivates us to combine visual evidence with original-image geometry for score refinement without changing the original scoring pipelines. In this paper, we introduce MetaRefine, a metadata-aware score refinement architecture whose lightweight residual heads predict score corrections from geometry, multi-level visual features, and the frozen model's original score. Across 100 configurations on 21 datasets, MetaRefine improves PLCC in 88 configurations and SRCC in 85, with average relative gains of 7.55% and 7.10%, respectively. It reduces MAE in 90 configurations and RMSE in 89, with average relative reductions of 23.27% and 21.93%, respectively. The median parameter overhead is only 1.11%. MetaRefine can further improve state-of-the-art (SOTA) assessment models on multiple benchmarks.

Figures & tables

Figure 2: Refinement across image geometries and input pipelines. (A) Original and refined MAE across aspect-ratio bins for six assessment models [ 52 , 2 , 16 , 8 , 18 ] . (B) Illustrative score corrections under the models’ original cropping, resizing, padding, patch sampling, and native-resolution inputs [ 52 , 44 , 24 , 16 , 21 ] . Red/green: original/refined scores. Blue parentheses: ground truth.

Figure 3: Overview of MetaRefine. Metadata-aware score refinement using an existing frozen assessment model, with the residual head detailed at the bottom right. Only the residual heads are optimized.

Figure 4: Visual examples of score refinement across datasets. Original (red), MetaRefine (green), and GT (blue).

KonIQ-10k (IQA) [ 22 ] - Test/Total: 2,015 / 10,073
	SRCC $\uparrow$			PLCC $\uparrow$			MAE $\downarrow$			RMSE $\downarrow$
Assessment model	Base	Ref.	Gain	Base	Ref.	Gain	Base	Ref.	Gain	Base	Ref.	Gain	Params $\uparrow$
DP-IQA (Student) [ 14 ]	0.9752	0.9781	+0.30%	0.9808	0.9828	+0.21%	0.0312	0.0203	+34.93%	0.0396	0.0285	+27.97%	0.26%
MANIQA [ 61 ]	0.9022	0.9505	+5.35%	0.9280	0.9639	+3.87%	0.1521	0.0292	+80.79%	0.1614	0.0377	+76.63%	0.71%
RALI [ 69 ]	0.9248	0.9440	+2.08%	0.9410	0.9552	+1.51%	0.0494	0.0354	+28.36%	0.0632	0.0458	+27.50%	0.24%
TReS [ 16 ]	0.9362	0.9406	+0.48%	0.9461	0.9554	+0.99%	0.1602	0.0348	+78.29%	0.1736	0.0465	+73.20%	0.74%

Table 1: Complete score-refinement results, part 1 of 2 (52 configurations). Base: frozen baseline assessment model. Ref.: the baseline model after MetaRefine refinement. Gain is the relative change from Base to Ref., computed by dividing the increase in SRCC/PLCC or the reduction in MAE/RMSE by the absolute Base value and expressing the result as a percentage. Params

\uparrow

: added trainable parameters relative to the complete frozen baseline (%). Darker green / red indicates larger relative improvement/decline, respectively, with color intensity capped at a 30% change. Improvements correspond to higher SRCC/PLCC or lower MAE/RMSE. Rows are sorted by refined SRCC within each dataset. Test/Total gives the default evaluation count and total sample count used in that dataset protocol. Data partitions and evaluation protocols are detailed in Appendices C.2 - C.4 .

	SRCC $\uparrow$			PLCC $\uparrow$			MAE $\downarrow$			RMSE $\downarrow$
Assessment model	Base	Ref.	Gain	Base	Ref.	Gain	Base	Ref.	Gain	Base	Ref.	Gain	Params $\uparrow$
AVA (IAA) [ 39 ] - Test/Total: 19,929 / 255,508
FGAesQ [ 63 ]	0.8720	0.8838	+1.35%	0.8676	0.8791	+1.32%	0.2885	0.2719	+5.75%	0.3785	0.3593	+5.07%	0.53%
ELTA [ 36 ]	0.8420	0.8569	+1.77%	0.8437	0.8581	+1.71%	0.3110	0.2959	+4.87%	0.4037	0.3864	+4.27%	0.62%
QPT V2 (IAA) r [ 58 ]	0.8551	0.8555	+0.05%	0.8649	0.8658	+0.11%	0.0307	0.0299	+2.54%	0.0389	0.0381	+2.03%	1.08%
ArtiMuse [ 5 ]	0.8265	0.8261	-0.05%	0.8250	0.8249	-0.01%	0.3265	0.3263	+0.07%	0.4249	0.4249	-0.00%	0.01%

Table 2: Complete score-refinement results, part 2 of 2 (48 configurations). Base: frozen baseline assessment model. Ref.: the baseline model after MetaRefine refinement. Gain is the relative change from Base to Ref., computed by dividing the increase in SRCC/PLCC or the reduction in MAE/RMSE by the absolute Base value and expressing the result as a percentage. Params

\uparrow

Reference Configuration
Variant	PLCC	SRCC	MAE	RMSE
Full	0.000%	0.000%	0.000%	0.000%
(A) Multi-Level Visual Features
w/o visual features	-3.638%	-3.889%	-55.656%	-48.445%
Final features only	-0.960%	-1.265%	-5.371%	-5.436%
w/o final features	-0.285%	+0.267%	-7.129%	-6.751%

Table 3: Ablation study result. Full denotes the complete MetaRefine configuration. Relative improvements over Full are averaged equally. Positive values indicate higher correlation or lower error. Red / green denotes negative/positive changes.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Train	Val.	Eval.	Configuration-specific settings
KonIQ-10k	7,058	1,000	2,015	QCN and MANIQA: 8,058 training images, MANIQA uses an alternative author list.
KADID-10k	8,000	–	2,125	NLNet: author content split, 6,000/2,000/2,125, 17 evaluation contents.
SPAQ	7,910	1,086	2,129	QualiCLIP+ and NIMA IR-v2: 9,013/–/2,112, QCN: 8,996/–/2,129.
LIVE Challenge	814	117	231	HyperIQA/HyperIQA+NT: 930/–/232, QualiCLIP+: author list, 931/–/231.
FLIVE	121,080	30,856	7,304	Default: whole images and cropped regions, DP-IQA and PaQ-2-PiQ: whole images only, 30,270/7,714/1,826.
AVA	235,579	–	19,929	AesCLIP: exposure-separated 248,259/–/7,249.

Appendix

Table S1 : Data partitions for the paired score-refinement comparisons. Counts denote samples in the designated refinement training, validation, and evaluation partitions, including whole images and cropped regions for FLIVE. A dash means no separate validation partition is listed. Configuration-specific exceptions appear in the last column, and model reproduction and adaptation settings are detailed in Appendix C.3 .

(A) Multi-Level Visual Features
Variant	PLCC $\uparrow$	SRCC $\uparrow$	MAE $\uparrow$	RMSE $\uparrow$
w/o visual features	-3.638% (83)	-3.889% (80)	-55.656% (82)	-48.445% (81)
Final features only	-0.960% (75)	-1.265% (73)	-5.371% (73)	-5.436% (77)
w/o final features	-0.285% (63)	+0.267% (61)	-7.129% (58)	-6.751% (61)
w/o early features	-0.261% (58)	-0.288% (57)	-0.097% (60)	-0.177% (61)
w/o middle features	-0.309% (64)	-0.353% (62)	-0.174% (58)	-0.221% (60)

Appendix

Table S2 : Component comparisons. Three-head variants versus the complete MetaRefine configuration (Full). In the upper panels, each cell gives the equally weighted mean relative improvement over Full and the number of configurations where Full is better. The final panel reports median changes for the three original-score pathway variants. Positive values indicate higher correlation or lower error. Negative/positive changes have red/green backgrounds.

Model (Dataset)	PLCC $\uparrow$		SRCC $\uparrow$		MAE $\downarrow$		RMSE $\downarrow$
Variant	Mean $\pm$ SD	Change	Mean $\pm$ SD	Change	Mean $\pm$ SD	Change	Mean $\pm$ SD	Change
TReS (FLIVE)
Frozen assessment model	0.6398		0.6293		0.1317		0.1470
Full	0.7468 ± 0.0008	0.00%	0.7104 ± 0.0007	0.00%	0.0444 ± 0.0001	0.00%	0.0588 ± 0.0001	0.00%
Final features only	0.7221 ± 0.0006	-3.30%	0.6804 ± 0.0011	-4.21%	0.0463 ± 0.0001	-4.28%	0.0611 ± 0.0001	-3.95%
Without final features	0.7424 ± 0.0012	-0.58%	0.7079 ± 0.0015	-0.34%	0.0450 ± 0.0001	-1.29%	0.0595 ± 0.0001	-1.19%

Appendix

Table S3 : Component ablations, part 1 of 2. Full denotes MetaRefine. Single-head means

\pm

sample standard deviations use seed group A (Appendix D.1 ). Changes compare each mean with the Full mean over the same seeds. Positive/green denotes improvement and negative/red decline.

MUSIQ (SPAQ)
Model (Dataset)	PLCC $\uparrow$		SRCC $\uparrow$		MAE $\downarrow$		RMSE $\downarrow$
Variant	Mean $\pm$ SD	Change	Mean $\pm$ SD	Change	Mean $\pm$ SD	Change	Mean $\pm$ SD	Change
Frozen assessment model	0.6954		0.6984		0.1824		0.2210
Full	0.9052 ± 0.0005	0.00%	0.9005 ± 0.0005	0.00%	0.0696 ± 0.0001	0.00%	0.0893 ± 0.0002	0.00%
Final features only	0.8690 ± 0.0007	-4.01%	0.8674 ± 0.0007	-3.68%	0.0818 ± 0.0003	-17.49%	0.1060 ± 0.0003	-18.69%
Without final features	0.8921 ± 0.0003	-1.45%	0.8887 ± 0.0001	-1.31%	0.0746 ± 0.0003	-7.23%	0.0958 ± 0.0003	-7.30%

Appendix

Table S4 : Component ablations, part 2 of 2. Full denotes MetaRefine. Single-head means

\pm

sample standard deviations use seed group A (Appendix D.1 ). Changes compare each mean with the Full mean over the same seeds. Positive/green denotes improvement and negative/red decline.

TReS (FLIVE)
Model (Dataset)	PLCC $\uparrow$		SRCC $\uparrow$		MAE $\downarrow$		RMSE $\downarrow$		Added
Heads	Score	Change	Score	Change	Score	Change	Score	Change	params
Base	0.6398		0.6293		0.1317		0.1470		0
1	0.7468 ± 0.0008	-0.05%	0.7104 ± 0.0007	-0.07%	0.0444 ± 0.0001	-0.08%	0.0588 ± 0.0001	-0.09%	377,153
3 (Full)	0.7472	0.00%	0.7109	0.00%	0.0444	0.00%	0.0588	0.00%	1,131,459
5	0.7471	-0.01%	0.7107	-0.02%	0.0444	0.00%	0.0588	-0.01%	1,885,765

Appendix

Table S5 : Single-head variability and residual-head ensembles. All refined rows use MetaRefine. One-head metrics are means

\pm

sample standard deviations over five seeds. Ensembles average predictions. Changes are relative to the default three-head Full. Added parameters count residual heads. The frozen model is shared.

TReS (FLIVE)
Variant	PLCC $\uparrow$		SRCC $\uparrow$		MAE $\downarrow$		RMSE $\downarrow$		Added params
Linear, unbounded	0.7590 ± 0.0000	+1.70%	0.7320 ± 0.0000	+3.18%	0.1251 ± 0.0000	-181.50%	0.1377 ± 0.0000	-134.04%	5,892
MLP (32), unbounded	0.7387 ± 0.0016	-1.03%	0.7025 ± 0.0020	-0.97%	0.0454 ± 0.0001	-2.08%	0.0599 ± 0.0002	-1.84%	188,577
MLP (64), Full	0.7463 ± 0.0010	0.00%	0.7094 ± 0.0016	0.00%	0.0444 ± 0.0000	0.00%	0.0588 ± 0.0001	0.00%	377,153
MLP (128), unbounded	0.7479 ± 0.0062	+0.21%	0.7096 ± 0.0085	+0.02%	0.0442 ± 0.0005	+0.44%	0.0587 ± 0.0007	+0.23%	754,305
MLP (64–64), unbounded	0.7398 ± 0.0016	-0.87%	0.7129 ± 0.0010	+0.49%	0.0531 ± 0.0010	-19.50%	0.0686 ± 0.0011	-16.55%	381,313

Appendix

Table S6 : Residual-head architecture, part 1 of 2. Single-head means

\pm

sample standard deviations over ten seeds. Changes are relative to the Full mean over the same seeds. Parameters are for one head. Full is the unbounded 64-unit MLP. All rows use Huber regression.

MUSIQ (SPAQ)
Variant	PLCC $\uparrow$		SRCC $\uparrow$		MAE $\downarrow$		RMSE $\downarrow$		Added params
Linear, unbounded	0.9007 ± 0.0000	-0.53%	0.8966 ± 0.0000	-0.46%	0.1642 ± 0.0000	-136.26%	0.1857 ± 0.0000	-107.99%	3,076
MLP (32), unbounded	0.9003 ± 0.0011	-0.57%	0.8952 ± 0.0010	-0.61%	0.0719 ± 0.0002	-3.47%	0.0926 ± 0.0003	-3.76%	98,465
MLP (64), Full	0.9054 ± 0.0004	0.00%	0.9008 ± 0.0005	0.00%	0.0695 ± 0.0002	0.00%	0.0893 ± 0.0002	0.00%	196,929
MLP (128), unbounded	0.9078 ± 0.0003	+0.26%	0.9035 ± 0.0002	+0.31%	0.0687 ± 0.0001	+1.16%	0.0880 ± 0.0001	+1.47%	393,857
MLP (64–64), unbounded	0.9054 ± 0.0005	-0.01%	0.9004 ± 0.0006	-0.04%	0.0727 ± 0.0010	-4.58%	0.0928 ± 0.0011	-3.90%	201,089

Appendix

Table S7 : Residual-head architecture, part 2 of 2. Single-head means

\pm

sample standard deviations over ten seeds. Changes are relative to the Full mean over the same seeds. Parameters are for one head. Full is the unbounded 64-unit MLP. All rows use Huber regression.

Bound	Regression	PLCC $\uparrow$	SRCC $\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	Full wins
0.1875	Huber	-2.25% (-0.55%)	-2.29% (-0.40%)	-54.97% (-8.01%)	-48.18% (-7.08%)	61
0.375	Huber	-1.38% (-0.30%)	-1.35% (-0.20%)	-45.27% (-3.46%)	-40.41% (-3.06%)	56
0.75	Huber	-0.65% (-0.17%)	-0.75% (-0.05%)	-31.74% (-1.02%)	-29.12% (-0.55%)	56
1.5	Huber	+0.48% (-0.03%)	+0.53% (-0.01%)	-16.76% (-0.08%)	-16.25% (-0.13%)	47
3.375	Huber	+2.04% (0.00%)	+1.73% (0.00%)	-7.13% (0.00%)	-7.51% (-0.01%)	35
Adaptive (90th percentile)	Huber	+2.02% (-0.03%)	+1.73% (-0.02%)	-7.44% (-0.06%)	-7.85% (-0.08%)	41

Appendix

Table S8 : Residual-bound comparison under the same evaluation settings as Table 3 of the main paper. Three-head results with Huber regression. Each cell gives the equally weighted mean (median) relative change from Full. Positive means improvement for every metric. Green/red shading indicates mean improvement/decline. Full wins counts configurations where Full outperforms the variant on all four metrics. Full is bold.

TReS (FLIVE)
Variant	PLCC $\uparrow$		SRCC $\uparrow$		MAE $\downarrow$		RMSE $\downarrow$
Heads	Change	95% interval	Change	95% interval	Change	95% interval	Change	95% interval
10	-0.06%	[-0.00068, -0.00015]	-0.14%	[-0.00132, -0.00068]	+0.05%	[-0.00001, +0.00005]	+0.01%	[-0.00002, +0.00004]
PIAA-MIR (Swin-T) (PARA GIAA)
10	-0.01%	[-0.00019, +0.00009]	+0.01%	[-0.00018, +0.00032]	-0.04%	[-0.00005, +0.00001]	-0.04%	[-0.00006, +0.00001]
CLIP-IQA+ (L/14) (KonIQ-10k)

Appendix

Table S9 : Paired uncertainty of the head-count comparison. Ten-head relative changes and 95% bootstrap confidence intervals for absolute changes from the default three-head Full, computed separately for each metric. Positive values mean improvement for every metric. Intervals use 2,000 paired resamples, without adjustment for multiple comparisons. Independent units: 1,826 reference-ID groups (TReS), 2,952 groups of images with identical pixels (PIAA-MIR), 2,015 groups of images with identical pixels (CLIP-IQA+), 2,114 groups of images with identical pixels (MUSIQ), and 19,929 images (TOPIQ).

Charm (AVA)
Heads	Median ms/image $\downarrow$	Images/s $\uparrow$	Allocated MiB $\downarrow$	Reserved MiB $\downarrow$
0	72.94	13.36	1223.93	1250.00
1	74.83	13.03	1231.95	1248.00
3	75.40	13.26	1235.96	1266.00
5	75.49	13.07	1239.96	1268.00
10	76.86	12.74	1249.97	1268.00

Appendix

Table S10 : End-to-end inference efficiency on an RTX 3090. Batch size 1, 24 timed image calls per setting.

K=0

denotes the frozen model without refinement. Memory is peak PyTorch allocated/reserved memory, including the assessment model and residual heads on the GPU. Green/red cells indicate lower/higher cost or higher/lower throughput relative to each model’s

K=0

. Saturation reflects the relative change.

Model (Dataset)	Train images	Fit (s)	Load (s)	Norm./GPU (s)	Cache GiB	Train feat. GiB	Extract/write 8 (s)
Charm (AVA)	235579	5.287	0.196	57.464	10.808	7.189	0.608/0.003
CLIP-IQA+ (L/14) (KonIQ-10k)	7058	0.843	1.147	0.278	0.205	0.141	0.669/0.003
NLNet (KADID-10k)	6000	0.963	0.357	0.051	0.056	0.031	6.682/0.004
MUSIQ (SPAQ)	7910	0.790	0.472	0.136	0.131	0.091	16.184/0.003
FGResQ (FGRestore)	14700	0.878	3.874	1.043	0.796	0.631	0.241/0.003

Appendix

Table S11 : Residual-head training and cache costs. One full-batch 85-update head fit, cache reading, and feature normalization with GPU transfer are timed separately. Cache size counts the stored files. Training-feature size counts only the visual feature values used by the full head, at four bytes per value.

Figure S1 : Qualitative score refinement on AVA. Original scores (red) and scores after MetaRefine (green) are shown together for each assessment model. Ground truth is blue. Scores use the AVA 1–10 scale.

Figure S2 : Qualitative score refinement on KonIQ-10k. Scores use the normalized 0–1 evaluation scale. Colors and paired-score notation follow Figure S1 .

Figure S3 : Qualitative score refinement on SPAQ. Scores use the normalized 0–1 evaluation scale. Each model retains its own original scoring pipeline. Colors and notation follow Figure S1 .

Figure S4 : Qualitative score refinement on KADID-10k. Examples span multiple contents and distortion types on the normalized 0–1 scale. Colors and notation follow Figure S1 .

Figure S5 : Qualitative score refinement on TAD66K. All scores are displayed on the normalized 0–1 scale. Charm’s original and refined scores are both divided by 10 to match this scale. Colors and notation follow Figure S1 .

Figure S6 : Qualitative score refinement on AADB. Scores use the dataset’s 0–1 scale. Colors and paired-score notation follow Figure S1 .

Image Quality Assessment

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

See all 175 papers

Q-DeepSight: Incentivizing Thinking with Images for Image Quality Assessment and Refinement
Apr 18, 2026Xudong Li, Jiaxi Tan, Ziyin Zhou +6Tool Use in VLMs Multimodal CoT Reasoning
Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, 361005, P.R. China · Beijing Institute of Technology
PortraitAes: Intent-Conditioned Structured Portrait Aesthetic Assessment
Oct 4, 2026Junzhou Xie, Haozhong Xiong, Xunyun Tian +8Image Quality Assessment
Southeast University · Qwen Business Unit of Alibaba
MSD-Score: Multi-Scale Distributional Scoring for Reference-Free Image Caption Evaluation
May 7, 2026Shichao Kan, Xuyang Zhang, Haojie Zhang +7Cross-Modal Alignment Vision-Language Grounding
School of Computer Science and Engineering, Central South University, Changsha, 410083, China · School of Computer Science and Technology and the State Key Laboratory of Advanced Rail Autonomous Operation, Beijing Jiaotong University, Beijing, 100044, China · School of Computer Science and Technology, University of Chinese Academy of Sciences, Beijing, 101408, China +1

Apr 18, 2026J/K move · Enter open · S save

Q-DeepSight: Incentivizing Thinking with Images for Image Quality Assessment and Refinement

Xudong Li, Jiaxi Tan, Ziyin Zhou, Yan Zhong+5

Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, 361005, P.R. China · Beijing Institute of Technology

Tool Use in VLMs Multimodal CoT Reasoning Image Quality Assessment RL for VLMs

Open paper AskPDF arXiv

Image Quality Assessment (IQA) models are increasingly deployed as perceptual critics to guide generative models and image restoration. This role demands not only accurate scores but also actionable, localized feedback. However, current MLLM-based methods adopt a single-look, language-only paradigm, which departs from human evidence-seeking judgment and yields weakly grounded rationales, limiting their reliability for in-the-loop refinement. We propose Q-DeepSight, a think-with-image framework that emulates this human-like process. It performs interleaved Multimodal Chain-of-Thought (iMCoT) with tool-augmented evidence acquisition (e.g., crop-and-zoom) to explicitly determine where quality degrades and why. To train these long iMCoT trajectories via reinforcement learning, we introduce two techniques: Perceptual Curriculum Reward (PCR) to mitigate reward sparsity and Evidence Gradient Filtering (EGF) to improve credit assignment for visually-grounded reasoning. Q-DeepSight achieves state-of-the-art performance across diverse benchmarks, including natural, restored, and AI-generated content. Furthermore, we demonstrate its practical value with Perceptual-in-Generation (PiG), a training-free framework where Q-DeepSight's diagnoses guide iterative image enhancement, effectively closing the loop between assessment and refinement.

Abstract

Image Quality Assessment

Explore similar work

Q-DeepSight: Incentivizing Thinking with Images for Image Quality Assessment and Refinement

PortraitAes: Intent-Conditioned Structured Portrait Aesthetic Assessment

MSD-Score: Multi-Scale Distributional Scoring for Reference-Free Image Caption Evaluation