An AI model designed to judge the quality of computer-generated images has reported strong agreement with benchmark ratings across five public datasets after treating image quality and prompt fit as connected parts of the same task. Called ACL-IQA, it uses one collaborative path and one adversarial path to model how those two judgments can interact.
The work is an arXiv version 1 preprint dated 25 Aug 2026. It is an offline test of predictions against existing benchmark ratings, not a new recruited-participant study. Its central premise is that perception quality and prompt alignment can sometimes reinforce each other and sometimes conflict.
A model built around a tension
ACL-IQA is built as a Dual-Gated Mixture-of-Experts framework. In plain terms, the design lets separate collaborative and adversarial routes contribute to the two predictions, with the authors interpreting that structure as an adaptive balance between reinforcement and disentanglement.
The main evaluation covered five public benchmark datasets: AGIQA-1K with 1,080 images, AGIQA-3K with 2,982, AIGCIQA2023 with 2,400, AGIQA-20K with 20,000 and EvalMi-50K with 50,400. The records contained average human ratings for perception and alignment; AIGCIQA2023 also included an authenticity task.
The evaluation used the AGIQA-20K repository train/test split, five-fold cross-validation for EvalMi-50K and 10 repeated runs with identical splits for the other datasets, with average results reported. Because the split protocols were not identical, the scores are best read as dataset-specific results rather than one perfectly uniform ranking.
The benchmark picture
On AGIQA-3K, the model’s perception scores were 0.8969 for SRCC, 0.9282 for PLCC and 0.9126 for the reported mean. For alignment, the corresponding figures were 0.7832, 0.8800 and 0.8316. SRCC and PLCC are two ways of measuring how closely model scores track benchmark ratings; the mean score is the paper’s overall summary.
On AGIQA-1K, perception SRCC was 0.8617, PLCC was 0.8885 and the mean score was 0.8751. AGIQA-20K produced SRCC of 0.8858, PLCC of 0.9113 and a mean score of 0.8986.
AIGCIQA2023 gave the model three reported tasks. Perception scores were 0.8720, 0.8865 and 0.8793; authenticity scores were 0.8180, 0.8083 and 0.8132; and alignment scores were 0.8092, 0.8005 and 0.8049, in SRCC, PLCC and mean-score order.
On EvalMi-50K, perception reached 0.8810, 0.9050 and 0.8930, while alignment reached 0.8664, 0.8795 and 0.8730, again in SRCC, PLCC and mean-score order.
Cross-dataset tests provided a check on whether the model would carry over between benchmark sets. From AIGCIQA2023 to AGIQA-3K, the reported SRCCs were 0.761 for perception and 0.676 for alignment, with stated relative improvements of 9.5% and 14.4% over IPCE. From AGIQA-3K to EvalMi-50K, ACL-IQA’s perception SRCC was 0.6057, compared with 0.2914 for MA-AGIQA.
Prompt-level analysis offered a second check on the proposed interaction. It reported adversarial prompts at 41.40% of 6,846 prompts in RichHF-18K, 36.56% of 2,090 prompts in EvalMi-50K, 33.33% in AGIQA-3K and 24.00% in AIGCIQA2023. Cosine-similarity distributions covered both positive and negative ranges.
Where the evidence stops
In an AGIQA-3K ablation, the complete model reported perception SRCC and PLCC of 0.8903 and 0.9282, and alignment SRCC and PLCC of 0.7832 and 0.8800. The corresponding values for the ablated experiment were 0.8841 and 0.9246 for perception, and 0.7697 and 0.8725 for alignment. The table reports point metrics without uncertainty intervals.
Supplementary analysis used 10-fold cross-validation on AGIQA-3K and AIGCIQA2023, with 95% confidence intervals and paired t-tests comparing ACL-IQA with IPCE. The authors report p-values well below 0.01 across all evaluations. Those formal interval and significance analyses were limited to the two datasets.
Those results do not show that ACL-IQA causes people to make better judgments. The study predicts agreement with existing mean opinion scores in public benchmark data; it does not report a new human-rating experiment or a deployment outcome.
The interaction claim also remains provisional. The paper does not report a human-rater sample or a numerical estimate of interaction size, and the prompt proportions are dataset- and network-level diagnostics rather than participant-level estimates. Independent validation with newly collected, prespecified human ratings remains an open question.
Direct comparisons need care because the datasets use different split protocols, the main benchmark tables generally lack confidence intervals and p-values, and formal significance analysis is limited to AGIQA-3K and AIGCIQA2023. The preprint says its code is available at the listed GitHub repository and reports support from the National Natural Science Foundation of China and the Guangdong Philosophy and Social Sciences Planning Project.
Taken together, the paper presents ACL-IQA as a benchmarked modeling proposal for combining perception and prompt alignment. Its strongest evidence is agreement with existing labels across several datasets; whether the interaction pattern holds for unseen generators, prompts and rating populations remains unresolved.
Paper data and sources
Original title: Bridging Adversarial and Collaborative Learning for AI-Generated Image Quality Assessment
Authors: Baoliang Chen, Qing Lin, Sijie Mai
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text