Bridging Adversarial and Collaborative Learning for AI-Generated Image Quality Assessment

Baoliang Chen; Qing Lin; Sijie Mai

cs.CVAug 25, 2026

Bridging Adversarial and Collaborative Learning for AI-Generated Image Quality Assessment

Authors: Baoliang Chen, Qing Lin, Sijie Mai

Organizations: Department of Computer Science, South China Normal University, China · Department of Data Science and Analytics, Hong Kong University of Science and Technology (Guangzhou)

Abstract

AI-generated image quality assessment (AIGIQA) requires jointly reasoning about perceptual fidelity and prompt alignment, two quality dimensions that are often treated as independent in existing AIGIQA models. However, by re-examining human ratings, we uncover a previously overlooked phenomenon: the two dimensions are interdependent and exhibit both competitive and cooperative interactions during human rating. This observation suggests that a unified model should neither collapse the two dimensions nor rigidly separate them, but rather adaptively negotiate their interplay. Motivated by this insight, we introduce an interaction-aware learning framework that models perception-alignment relations through adversarial and collaborative inference pathways. Instead of designing a rigid dual-branch architecture, our method employs a gated interaction module that dynamically routes features according to the inferred relationship between the two dimensions. Task-aware prompts further modulate the gating behaviour, enabling the model to switch between competition and cooperation when necessary. Experiments across multiple AIGIQA benchmarks demonstrate that our approach not only achieves state-of-the-art accuracy but also yields interpretable interaction patterns, offering a more faithful approximation of human judgment. The codes are available at https://github.com/LQAMEI/ACL-IQA.

Figures & tables

Fig. 1: Illustration of adversarial and collaborative interactions in human rating during AIGIQA. The ground-truth quality scores of each dimension are depicted in the line charts.

Fig. 2: Overview of our AIGIQA model incorporating both adversarial and collaborative learning.

Symbol	Definition
$T_{col},T_{adv}$	Textual instructions for collaborative and adversarial learning strategy.
$T_{per},T_{aln}$	Textual prompt templates for perception and alignment assessment.
$F_{col}^{T},F_{adv}^{T}$	Textual learning instruction features derived from templates $T_{col}$ and $T_{adv}$ .
$F_{per}^{T},F_{aln}^{T}$	Textual quality assessment features derived from templates $T_{per}$ and $T_{aln}$ .
$F_{col}^{I},F_{adv}^{I}$	Decoupled image features routed for collaborative and adversarial learning paths.
$G_{c}(\cdot),G_{a}(\cdot)$	Dual gating networks guided by instructional textual features.

TABLE I: Illustration of key mathematical symbols and notations.

Fig. 3: Design details of our proposed dual-gated MoE block.

Fig. 4: Illustration of five types of convolution kernels [ 5 ] . The arrow represents the difference between the weight at its starting point and the weight at its endpoint.

Method	SRCC	PLCC	Mean Score
CNNIQA [ 57 ]	0.5800	0.7139	0.6470
DBCNN [ 58 ]	0.7491	0.8211	0.7851
MGQA [ 59 ]	0.6011	0.6760	0.6386
CLIP-IQA [ 60 ]	0.8227	0.8411	0.8319
StairIQA [ 61 ]	0.5504	0.6088	0.5796
PSCR [ 62 ]	0.8430	0.8403	0.8417

TABLE II: Performance comparison of perception quality prediction on the AGIQA-1K. The best two results are in boldface.

Perception Quality				Alignment Quality
Method	SRCC	PLCC	Mean Score	Method	SRCC	PLCC	Mean Score
CNNIQA [ 57 ]	0.7478	0.8469	0.7974	CLIP [ 12 ]	0.5972	0.6839	0.6406
DBCNN [ 58 ]	0.8207	0.8759	0.8483	ImageReward [ 63 ]	0.7298	0.7862	0.7580
CLIP-IQA [ 60 ]	0.8426	0.8053	0.8240	HPS [ 64 ]	0.6623	0.7008	0.6816
ManIQA [ 31 ]	0.8767	0.9105	0.8936	ManIQA [ 31 ]	–	–	–
Re-IQA [ 32 ]	0.8424	0.8890	0.8657	Re-IQA [ 32 ]	–	–	–

TABLE III: Performance comparison of perception and alignment quality prediction on the AGIQA-3K dataset. The best two results in each column are highlighted in boldface. * denotes methods evaluated by directly loading pretrained model weights for inference.

Method	Perception			Authenticity			Alignment
Method	SRCC	PLCC	Mean Score	SRCC	PLCC	Mean Score	SRCC	PLCC	Mean Score
CNNIQA [ 57 ]	0.7160	0.7937	0.7549	0.5958	0.5734	0.5846	0.4758	0.4937	0.4848
ResNet34 [ 67 ]	0.7229	0.7578	0.7404	0.5998	0.6285	0.6142	0.7058	0.7153	0.7106
TReS [ 30 ]	0.8064	0.8347	0.8206	–	–	–	–	–	–
ManIQA [ 31 ]	0.8269	0.8464	0.8367	–	–	–	–	–	–
Re-IQA [ 32 ]	0.8372	0.8721	0.8547	–	–	–	–	–	–

TABLE IV: Performance comparison of perception, authenticity, and alignment quality prediction on the AIGCIQA2023 dataset. The best two results in each column are highlighted in boldface. * denotes methods evaluated by directly loading pretrained model weights for inference.

Method	SRCC	PLCC	Mean Score
CNNIQA [ 57 ]	0.5968	0.5913	0.5941
DBCNN [ 58 ]	0.8506	0.8688	0.8597
HyperIQA [ 2 ]	0.8160	0.8320	0.8240
MUSIQ [ 68 ]	0.8320	0.8640	0.8480
ManIQA [ 31 ]	0.8500	0.8870	0.8685
StairIQA [ 61 ]	0.7890	0.8420	0.8155

TABLE V: Performance comparison on the AGIQA-20K dataset. The best two results in each column are highlighted in boldface. * denotes methods evaluated by directly loading pretrained model weights for inference.

Method	Perception Quality			Alignment Quality
Method	SRCC	PLCC	Mean Score	SRCC	PLCC	Mean Score
CNNIQA [ 57 ]	0.4348	0.5583	0.4966	0.1186	0.0791	0.0989
DBCNN [ 58 ]	0.5525	0.6181	0.5853	0.3301	0.3515	0.3408
HyperIQA [ 2 ]	0.5872	0.6768	0.6320	0.5348	0.5447	0.5398
TReS [ 30 ]	0.3935	0.4301	0.4118	0.1406	0.1520	0.1463
MUSIQ [ 68 ]	0.7985	0.8379	0.8182	0.5310	0.5510	0.5410

TABLE VI: Performance comparison on the EvalMi-50K dataset across perception and alignment dimensions. The best two results in each column are highlighted in boldface.

Methods	AIGCIQA2023 ${}^{\text{Train}}\rightarrow$ AGIQA-3K ${}^{\text{Test}}$			AGIQA-3K ${}^{\text{Train}}\rightarrow$ AIGCIQA2023 ${}^{\text{Test}}$
Methods	Perception	Alignment	Mean Score	Perception	Alignment	Mean Score
ResNet50 [ 67 ]	0.576	0.473	0.525	0.599	0.432	0.516
ViT-B/32 [ 13 ]	0.509	0.434	0.472	0.517	0.381	0.449
DB-CNN [ 58 ]	0.627	0.390	0.509	0.654	0.470	0.562
HyperIQA [ 2 ]	0.657	0.418	0.538	0.669	0.464	0.567
TIER [ 7 ]	0.611	0.417	0.514	0.675	0.441	0.558

TABLE VII: Cross-dataset prediction performance comparison. The best two results in each column are highlighted in boldface.

Method	AGIQA-3K ${}^{\text{Train}}\rightarrow$ EvalMi-50K ${}^{\text{Test}}$			AIGCIQA2023 ${}^{\text{Train}}\rightarrow$ EvalMi-50K ${}^{\text{Test}}$
Method	Perception	Alignment	Mean Score	Perception	Alignment	Mean Score
ManIQA [ 31 ]	0.0619	–	0.0619	0.1056	–	0.1056
Re-IQA [ 32 ]	0.1791	–	0.1791	0.1528	–	0.1528
TReS [ 30 ]	-0.0677	–	-0.0677	-0.1296	–	-0.1296
MA-AGIQA [ 42 ]	0.2914	–	0.2914	0.4438	–	0.4438
LIQE [ 6 ]	0.1823	0.1857	0.1840	0.1805	0.1568	0.1687

TABLE VIII: Cross-dataset performance comparison when transferring from early datasets to the EvalMi-50K. The best two results in each column are highlighted in boldface.

Exp. ID	MoE	Dual LS	Adv.	Col.	AGIQA-3K				AIGCIQA2023
					Perception		Alignment		Perception		Alignment
					SRCC	PLCC	SRCC	PLCC	SRCC	PLCC	SRCC	PLCC
1	✗	✗	✗	✗	0.8841	0.9246	0.7697	0.8725	0.8640	0.8788	0.7979	0.7887
2	✗	✗	$\checkmark$	$\checkmark$	0.8866	0.9253	0.7766	0.8771	0.8677	0.8835	0.8032	0.7939
3	$\checkmark$	✗	$\checkmark$	$\checkmark$	0.8895	0.9270	0.7787	0.8780	0.8687	0.8839	0.8043	0.7950
4	$\checkmark$	$\checkmark$	$\checkmark$	✗	0.8680	0.8682	0.7660	0.8675	0.8622	0.8787	0.7648	0.7545

TABLE IX: Ablation study results on the AGIQA-3K and AIGCIQA2023 datasets. “Dual LS” denotes the dual (adversarial and collaborative) learning text/instructions. “Adv.” and “Col.” refer to the adversarial and collaborative learning paths, respectively.

Fig. 5: Attention visualization of features from the MoE layer. The first and second rows correspond to the perceptual quality and alignment prediction tasks, respectively. “Adversarial” and “Collaborative” indicate that the features are learned from the two distinct interaction paths.

Fig. 6: Performance comparison under extreme scenarios, collaborative scenarios (left) where perception quality and alignment quality are positively correlated, and adversarial scenarios (right) where they conflict. Each row displays five generated images alongside their respective prompts. The bottom charts present quantitative comparisons across multiple image sets with perception and alignment quality scores.

Fig. 7: gMAD [ 72 ] comparison between ACL-IQA and IPCE on perception (top) and alignment (bottom) quality. (a) IPCE predicts both images as low quality, (b) IPCE predicts both images as high quality, (c) both images have low MOS, and (d) both images have high MOS. ACL-IQA aligns more consistently with human judgments across all cases.

Fig. 8: Comparison of gating activations in dual path learning. “Adv.” and “Col.” denote the adversarial and collaborative learning paths, respectively.

Fig. 9: The cosine similarity between the text feature and the image features ( $F_{col}^{I}$ , $F_{adv}^{I}$ ) extracted in collaborative and adversarial learning.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Method	FT	EvalMuse	GenAIBench	RichHF
LIQE [ 6 ]	$\checkmark$	0.7708 / 0.7583	0.5528 / 0.5513	0.6884 / 0.6616
FGA-BLIP2 [ 1 ]	$\checkmark$	0.7723 / 0.7716	0.5639 / 0.5578	0.6548 / 0.6415
Q-Align [ 7 ]	✗	0.3236 / 0.3241	0.2539 / 0.2275	0.2654 / 0.2774
Q-Align [ 7 ]	$\checkmark$	0.7467 / 0.7375	0.6536 / 0.6365	0.7120 / 0.7071
LMM4LMM [ 8 ]	✗	0.7284 / 0.6638	0.7445 / 0.5870	0.7193 / 0.6589
LMM4LMM [ 8 ]	$\checkmark$	0.8102 / 0.8038	0.7459 / 0.7244	0.7485 / 0.7293

Appendix

TABLE B.1: Performance comparison measured by SRCC and PLCC on EvalMuse-40K, GenAIBench, and RichHF-18K focusing on Alignment quality. The FT column indicates whether the model was fine-tuned. Both the best and second-best results are highlighted in bold.

Fig. B.1: t-SNE visualization of textual prompt embeddings extracted via Qwen3-Embedding-0.6B [ 11 ] , demonstrating the expanded semantic manifold of contemporary benchmarks.

Fig. B.2: t-SNE visualization of generated image features extracted via CLIP ViT-B/32 [ 12 ] , illustrating the pronounced visual distribution shift in modern generators.

Fig. C.1: Prompt-level Semantic Quadrant Analysis across AGIQA-3K, AIGCIQA2023, RichHF-18K, and EvalMi-50K datasets. Each point represents the mean perception and alignment scores of all images generated under a specific prompt by various models. The distribution explicitly delineates collaborative scenarios in the first and third quadrants and adversarial scenarios in the second and fourth quadrants. The regression line and shaded band represent the global positive data trend and its 95% confidence interval, while intersecting lines denote dataset medians.

Fig. C.2: Network-level Gradient Conflict Tracking. The histograms display the distribution of batch-wise cosine similarities between the alignment gradients and the perception or artifact gradients during training across the AGIQA-3K, AIGCIQA2023, RichHF-18K, and EvalMi-50K datasets. The curve indicates the Kernel Density Estimation, and the vertical line marks the zero-similarity boundary.

Fig. D.1: Feature maps from convolution experts for perception and alignment quality prediction. Each section displays pixel-level and patch-level features. Red boxes denote the top-3 dynamically gated experts, which are fused to form the final map, involving Vanilla Convolution, Central Difference Convolution, Angular Difference Convolution, Radial Difference Convolution, Horizontal Difference Convolution, and Vertical Difference Convolution.

Dataset	Task	Model	SRCC 95% CI	PLCC 95% CI	Paired $t$ -test ( $p$ -value)
AGIQA-3K	Perception	IPCE	$[0.8775,0.8907]$	$[0.9213,0.9279]$	$p<0.01$ (SRCC: 0.0031, PLCC: 0.0040)
	Perception	ACL-IQA	$\mathbf{[0.8840,0.8966]}$	$\mathbf{[0.9249,0.9316]}$	$p<0.01$ (SRCC: 0.0031, PLCC: 0.0040)
	Alignment	IPCE	$[0.7563,0.7831]$	$[0.8624,0.8826]$	$p<0.01$ (SRCC: 0.0069, PLCC: 0.0018)
	Alignment	ACL-IQA	$\mathbf{[0.7694,0.7969]}$	$\mathbf{[0.8698,0.8902]}$	$p<0.01$ (SRCC: 0.0069, PLCC: 0.0018)
AIGCIQA2023	Perception	IPCE	$[0.8562,0.8718]$	$[0.8736,0.8840]$	$p<0.01$ (SRCC: 0.0094, PLCC: 0.0013)
	Perception	ACL-IQA	$\mathbf{[0.8650,0.8790]}$	$\mathbf{[0.8806,0.8925]}$	$p<0.01$ (SRCC: 0.0094, PLCC: 0.0013)

Appendix

TABLE E.1: Statistical Significance Analysis using 10-Fold Cross-Validation comparing ACL-IQA against the competitive baseline IPCE. The

p

-values are computed using a paired

t

-test, where a value less than 0.01 indicates an extremely significant improvement.

Prompt Template	Perception	Alignment
Default Template	0.8719 / 0.8982	0.8556 / 0.8673
Concise Template	0.8705 / 0.8985	0.8545 / 0.8660
Verbose Template	0.8719 / 0.8991	0.8551 / 0.8672
Synonym Template	0.8693 / 0.8978	0.8531 / 0.8639

Appendix

TABLE E.2: Performance comparison using SRCC and PLCC of different prompt templates when trained on the EvalMi-50K dataset.

Dataset	$F_{col}$	$F_{adv}$	SRCC for Perception	PLCC for Perception	SRCC for Alignment	PLCC for Alignment
AGIQA-3K	$\checkmark$	$\checkmark$	0.8969	0.9162	-	-
	$\checkmark$		0.8819	0.8858	0.7207	0.8125
		$\checkmark$	0.8282	0.7777	0.3757 ( $\downarrow$ 47.9%)	0.0299 ( $\downarrow$ 96.3%)
AIGCIQA2023	$\checkmark$	$\checkmark$	0.8720	0.8865	-	-
	$\checkmark$		0.8379	0.8588	0.7564	0.7508
		$\checkmark$	0.8403	0.8576	-0.0427 ( $\downarrow$ 105.6%)	0.0438 ( $\downarrow$ 94.2%)

Appendix

TABLE F.1: Cross task feature decoupling evaluation for the ACL-IQA model trained to predict Perceptual Quality.

F_{col}

denotes the Collaborative Feature, and

F_{adv}

denotes the Adversarial Feature. The symbol

\checkmark

indicates the active feature used for prediction, while the downward arrow indicates the relative performance decrease compared to the collaborative path.

Dataset	$F_{col}$	$F_{adv}$	SRCC for Alignment	PLCC for Alignment	SRCC for Perception	PLCC for Perception
AGIQA-3K	$\checkmark$	$\checkmark$	0.7832	0.8800	-	-
	$\checkmark$		0.7161	0.8116	0.8788	0.8971
		$\checkmark$	0.7336	0.8248	0.2194 ( $\downarrow$ 75.0%)	0.2393 ( $\downarrow$ 73.3%)
AIGCIQA2023	$\checkmark$	$\checkmark$	0.8092	0.8005	-	-
	$\checkmark$		0.7257	0.7213	0.8469	0.8731
		$\checkmark$	0.7653	0.7655	0.4242 ( $\downarrow$ 49.9%)	0.3691 ( $\downarrow$ 57.7%)

Appendix

TABLE F.2: Cross task feature decoupling evaluation for the ACL-IQA model trained to predict Alignment Quality. The table presents the joint performance alongside isolated feature predictions, where the downward arrow indicates the relative performance decrease compared to the collaborative path.

Fig. F.1: Qualitative diagnostic panels for Perceptual Quality Assessment. The top rows illustrate the collaborative dominance regime where standard prompts benefit from shared alignment features. The bottom rows illustrate the adversarial dominance regime where complex prompts trigger semantic interference in the collaborative path, necessitating active feature decoupling via the adversarial path. Our joint ACL-IQA model dynamically balances both paths to match ground truth ratings.

Fig. F.2: Qualitative diagnostic panels for Alignment Quality Assessment. The top rows show scenarios where visual clarity is essential to verify fine grained semantic details, allowing the collaborative path to excel. The bottom rows demonstrate cases involving intentional stylistic degradation, where the adversarial path effectively strips away aesthetic bias to accurately evaluate text image correspondence.

Fig. F.3: Visualizations of implicit localization for artifacts and implausibility on RichHF-18K. The columns display the generated image, human-annotated Ground Truth mask, anomaly-targeted gradient heatmap from ACL-IQA, and textual prompt.

Fig. F.4: Visualizations of implicit localization for text-image misalignment on RichHF-18K. The columns display the generated image, human-annotated Ground Truth mask, anomaly-targeted gradient heatmap from ACL-IQA, and textual prompt.

Explore similar work

Jun 15, 2026

Decoupling Semantics from Distortions: Multi-Scale Two-Stream Vision-Language Alignment for AI-Generated Image Quality Assessment

Existing vision-language model (VLM)-based AI-generated image quality assessment (AIGIQA) methods suffer from a fundamental semantic-distortion dimensional conflict: monolithic representations optimized for semantic discrimination inherently entangle compositional understanding with low-level perceptual sensitivity, rendering them blind to fine-grained quality degradations. We introduce MST-CLIPIQA, a multi-scale two-stream framework that achieves hierarchical vision-language alignment through explicit representational decoupling. Our architecture leverages dual CLIP encoders with complementary patch granularities: coarse-grained streams capture global semantic coherence while fine-grained streams preserve textural signatures and artifact patterns. An information bottleneck-inspired gated fusion mechanism performs adaptive cross-scale distillation, with optional cross-attention enabling prompt-anchored correspondence evaluation when generation prompts are available. Extensive experiments across five benchmarks establish new state-of-the-art results, achieving average improvements of 1.11 percent SRCC on quality and 2.35 percent SRCC on text-image correspondence prediction, while maintaining efficiency with only 0.8M trainable parameters. Our project is available at https://github.com/YMlinfeng/MST-CLIPIQA.

Zijie Meng

Peking University

VLM Evaluation Image Quality Assessment Vision-Language Alignment Image Generation Evaluation

Ask PDF

Jul 6, 2026

Patch Knowledge Transfer for Efficient AI-Generated Image Quality Assessment

With the rapid advancement of image generation technologies, perceptual quality assessment of AI-generated images has emerged as a crucial research direction in computer vision. The core challenge of this task lies in achieving efficient quality assessment for massive generated images. Current mainstream approaches exhibit two key limitations: 1) Methods employing complex feature extraction strategies, while improving performance, incur prohibitive computational costs that hinder real-time inference; 2) Simple image scaling-based solutions, despite their computational efficiency, demonstrate significantly inferior assessment accuracy. To address this critical issue, we propose Patch Knowledge Transfer (PKT), a knowledge distillation-based optimization framework that achieves synergistic optimization of visual representation capability and inference efficiency through an innovative multi-level knowledge transfer mechanism. Specifically, we design a dual-model architecture: a teacher model with local-global hybrid processing provides high-quality supervision signals, while a student model relying solely on global processing efficiently inherits the teacher's representation capacity through multi-level supervision. Extensive experiments conducted on 4 AIGIQA databases demonstrate that the PKT framework enables the student model to maintain performance comparable to the teacher while reducing computational costs by 67.7%. Furthermore, compared to existing methods, our approach achieves a superior balance between model efficiency and assessment accuracy.

Jiquan Yuan

School of Software and Microelectronics, Peking University, Beijing, China

Image Quality Assessment Image Generation Evaluation Knowledge Distillation

Ask PDF

Sep 19, 2026

An Evolutionary Agentic Approach for Open-ended Image Quality Perception

Generative models are rapidly expanding image quality assessment (IQA) beyond traditional fidelity factors to emerging dimensions such as physical plausibility and text-rendering correctness. However, existing IQA models rely on fixed definitions and heavy supervision, making them difficult to extend to open-ended perceptual dimensions. We identify holistic bias as an important limitation: when scoring an unseen dimension, models reuse generic quality priors, leading to scoring errors and rank inversion. To address this, we propose PACE (Perceptual Agentic Collaborative Evolution), a training-free multi-agent framework that formulates open-ended IQA as explicit protocol construction. Given a target dimension, PACE uses collaborative agents to construct an evaluation protocol composed of verifiable Visual Question Answering (VQA) probes, grounding evaluation in concrete visual evidence rather than holistic impressions. The resulting protocol is calibrated using only four human-annotated images per dimension, while a dual-track scoring mechanism aligns model perception with human scoring scales. Across traditional IQA, structural fidelity, context-aware aesthetics, and newly defined open-ended dimensions, PACE consistently improves its MLLM backbone, achieving competitive performance across diverse IQA settings, and reduces the Holistic Override Rate (HOR) from 44.4% to 8.6%.

Zhenchen Tang, Bo Peng, Zichuan Wang +4

New Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing, China · School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China · The Hong Kong University of Science and Technology, Hong Kong SAR, China +1

Image Quality Assessment Vision-Language Model-as-a-Judge

Ask PDF