Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Q-Ponder: A Unified Training Pipeline for Reasoning-based Visual Quality Assessment

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Q-Ponder argues that accurate image-quality scoring and trustworthy reasoning descriptions are one joint objective, not a trade-off: a two-stage cold-start-plus-GRPO pipeline lets a 7B model score precisely and explain its judgment at the…

desk verdict The two-stage RL recipe and the score-regression results are genuinely good; the reasoning-description evaluation is too circular to support the unified claim. read the letter →

arxiv 2506.05384 v2 pith:WIFEHNJ2 submitted 2025-06-03 cs.CV cs.AIcs.MM

Zhuoxuan Cai , Jian Zhang , Xinbin Yuan , Peng-Tao Jiang , Wenxiang Chen , Bowen Tang , Lujian Yao , Qiyuan Wang
show 2 more authors
Jinwen Chen Bo Li
This is my paper
classification cs.CVcs.AIcs.MM
keywords imagequalityassessmentmultimodallargelanguagemodelreasoningdistillationGRPOreinforcementlearningcontinuousaccuracyrewardscoreregressionchain-of-thoughtdescriptionout-of-distributiongeneralization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that accurate quality score regression and interpretable reasoning descriptions are not competing goals for multimodal large language models in image quality assessment, and that a two-stage training recipe can deliver both at once. The recipe first distills expert-style chains of thought from a much larger teacher model into a 7B student, filtering out chains whose predicted scores disagree with human labels, then applies GRPO reinforcement learning with a continuous accuracy reward that couples the numerical score to the reasoning text. The resulting model, Q-Ponder, reports state-of-the-art score-regression accuracy on six IQA datasets, including a 6.5% relative SRCC gain over DeQA on the out-of-distribution CSIQ set, while also producing descriptions that score higher on completeness, accuracy, and reasonableness than description-focused baselines and even the 72B teacher. The potential payoff is a concrete mechanism by which explanation quality and regression precision reinforce each other, rather than forcing a choice between a black-box score and a fluent but loose description.

What carries the argument

The carrying mechanism is a two-stage training pipeline. In stage one, cold-start distillation, an expert-designed prompt asks a teacher model to assess low-level attributes (brightness, contrast, sharpness, noise, color, dynamic range, artifacts) and high-level attributes (subject clarity, composition, emotion, style, context), optionally with an enhanced distortion-tag hint such as 'severe blur'; a heuristic filter $D_{\text{refined}} = \{i \mid \operatorname{rank}(|\hat{y}_i - y_i|) \le \gamma N\}$ keeps only the chains whose predicted scores lie closest to the ground-truth scores, and the student is fine-tuned on the retained chains and true scores. In stage two, GRPO fine-tuning, the rewards are a dense accuracy reward $R_{\text{acc}} = \exp(-|\hat{y} - y|/\alpha)$ and a format reward $R_{\text{fmt}}$ for enclosing reasoning in '<think>...</think>' and the score in '<answer>...</answer>', combined as $R_{\text{total}} = \lambda_{\text{acc}} R_{\text{acc}} + \lambda_{\text{fmt}} R_{\text{fmt}}$. The dense accuracy reward is the coupling device: it gives the policy a smooth gradient on continuous quality labels and, the paper argues, simultaneously prunes hallucinated or inconsistent reasoning because those outputs tend to arrive at wrong scores.

What would settle it

Run a human-rated evaluation of Q-Ponder and DepictQA descriptions with the distortion tag withheld from both the models and the raters, then compare those ratings with the paper's judge-model scores; if Q-Ponder's accuracy and reasonableness lead over DepictQA collapses or reverses, the reasoning advantage is an artifact of tag-following rather than better quality perception.

Watch

Extended reading notes

Core claim

The paper's central claim is that score regression and reasoning generation can be jointly optimized in one multimodal large language model by placing the reward over the complete output, not over the score alone. On its evaluation protocol, Q-Ponder reports an average PLCC/SRCC of 0.845/0.821 across KonIQ, SPAQ, LIVE-Wild, KADID, CSIQ, and AGIQA, with CSIQ SRCC rising from 0.744 for DeQA to 0.792; on the reasoning-description benchmark, its average completeness/accuracy/reasonableness score is 4.182, versus 2.454 for DepictQA and 3.963 for the 72B teacher model. The authors attribute both gains to the same mechanism: expert reasoning chains distilled in the cold-start stage give the model a stable perceptual vocabulary, and the subsequent GRPO stage uses a dense accuracy reward to steer the policy toward outputs whose score is close to the human judgment and whose reasoning is consistent with that judgment. This is offered as evidence that the usual trade-off between numerical precision and interpretability in image quality assessment is not inherent.

Load-bearing premise

The load-bearing premise is that the reasoning-description evaluations are trustworthy: they are produced by a single proprietary MLLM judge that is shown the ground-truth distortion tag, the tested models are prompted with that same tag, and the paper's own Limitation states there is no human or second-model verification, so the reported reasoning gains may partly reflect hint-following rather than genuine perception.

Editorial extensions

If this is right

  • If the central claim holds, an MLLM for image quality assessment does not have to sacrifice explanations for score precision: one output can carry both a chain of thought and a regression value.
  • The continuous accuracy reward makes GRPO usable for continuous, subjective labels like MOS and DMOS, where binary correctness rewards would be almost always zero.
  • The ablation evidence implies that expert-length reasoning chains are load-bearing for score accuracy, not decoration: removing them or shortening them lowers PLCC/SRCC, especially on out-of-distribution sets.
  • The largest relative gains appear on cross-domain datasets, suggesting that structured reasoning acts as a regularizer that helps the model transfer across distortion types and image domains.
  • The same cold-start-then-RL recipe is readily transplantable to other score-plus-explanation vision-language tasks, such as aesthetic scoring or video quality assessment.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: because the evaluation judge and the tested models both see the ground-truth distortion tag, part of the reported reasoning advantage may come from echoing that hint rather than from perceptual diagnosis; a hint-free or human-judged rerun is needed to separate the two.
  • A testable extension the authors do not pursue is an anti-parroting reward term that penalizes verbatim repetition of the distortion tag, which would reveal whether the reasoning gains survive when copying is disincentivized.
  • The tolerance hyperparameter $\alpha = 10$ controls how sharply the policy is punished for score deviations; per-distortion or annealed schedules of $\alpha$ are a plausible way to push the out-of-distribution gains further.
  • The paper's joint-optimization claim is supported by ablations, but those ablations are correlational; a stronger causal test would corrupt the reasoning paths during training (for example, by swapping in wrong distortion labels) and check whether score regression degrades as a result.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes Q-Ponder, a two-stage training pipeline for multimodal large language model (MLLM) based image quality assessment. In the first stage, a 7B student model (Qwen2.5-VL-7B) is supervised fine-tuned on reasoning chains distilled from a 72B teacher using expert-designed prompts, with heuristic filtering based on prediction error and with ground-truth distortion tags supplied as hints. In the second stage, the student is fine-tuned with GRPO using a dense exponential accuracy reward and a binary format reward. The authors report state-of-the-art score regression on six IQA datasets (KonIQ, SPAQ, LIVE-Wild, KADID, CSIQ, AGIQA), with the largest gains on out-of-domain sets, and they report large improvements over description-based baselines on a three-dimension reasoning evaluation (completeness, accuracy, reasonableness) judged by Gemini-1.5-Pro. The paper also includes ablations showing that long chain-of-thought and reinforcement learning improve score regression over SFT-only variants.

Significance. The score-regression half of the paper is evaluated against external human-rated benchmarks, and the ablations in Table 3 support the internal claim that long CoT plus RL improves PLCC/SRCC beyond SFT-only training. If the score results reproduce, the paper offers a simple, reusable recipe for combining distillation and GRPO in IQA, and the smooth accuracy reward is a sensible alternative to sparse rewards for continuous quality labels. The reasoning-description half, however, is not established to the same standard: the evaluation in Sec. 3.1 and Appendix B is partly circular, the judge is a single proprietary model with no human or second-MLLM verification, and the severity labels are heuristic MOS bins rather than human distortion-severity annotations. Because the abstract's central claim is that Q-Ponder simultaneously achieves SOTA score regression and SOTA reasoning descriptions, the unvalidated reasoning protocol is a load-bearing weakness. The manuscript also provides no code, data, or evaluation scripts, which limits independent verification of both halves.

major comments (4)
  1. [Sec. 3.1, Appendix B, Sec. 4.2 (Table 2)] The reasoning 'Accuracy' metric is partly circular. In Appendix B the judge receives the ground-truth distortion tag and is asked whether the response identifies key distortions 'especially those relevant to the distortion tag' and whether severity aligns with the ground-truth label. In Appendix A the distillation prompt gives the model the same tag as a hint, and Sec. 4.2 states that all models are prompted with the distillation templates during evaluation. Q-Ponder was trained to incorporate this hint, whereas DepictQA, Co-Instruct, and Q-Instruct were not, so the large accuracy gap in Table 2 may largely measure whether a model repeats a provided label rather than whether it perceives distortions. Please re-run the reasoning evaluation without providing the distortion tag to either the tested models or the judge, or add a human-annotated distortion-identification benchmark, and report the results separately for the no-hint condition.
  2. [Sec. 4.4 (Limitation) and Table 4] The paper's own Limitation concedes that reasoning quality is judged by a single LLM without secondary verification from human experts or another proprietary MLLM. Since the abstract claims that Q-Ponder 'significantly outperforms description-based SOTA models' on description accuracy and reasonableness, this conceded lack of verification directly weakens the central claim. Please add human evaluation on a sample of responses and at least one additional judge model, and report inter-judge agreement or per-dimension score distributions rather than only means.
  3. [Sec. 3.1 and Table 2] The five-level severity labels are created by equal partitioning of the MOS range [0,1], not from human distortion-severity annotations. The Accuracy dimension therefore checks whether the model's severity statement matches a MOS-derived bin, which is a different construct from perceptual severity. This inflates the apparent accuracy gap and makes the severity component of the metric self-referential. Please either evaluate severity against human severity annotations or remove the severity sub-check and report the remaining distortion-identification accuracy separately.
  4. [Sec. 4.2 (Table 1)] The statement that 'All models are matched in parameter size and trained on the same 7k KonIQ subset for fairness' is not accurate as written, because handcrafted methods such as NIQE and BRISQUE and deep methods such as NIMA and MUSIQ are not trained on the 7k subset in the same sense as the MLLM baselines. Please clarify that the parameter-matching and training-subset statement applies only to the MLLM-based methods, and specify the exact training data and protocol used for Q-Insight, since its inclusion in the comparison affects the Table 1 ranking.
minor comments (5)
  1. [Sec. 3.3 (Eq. 4)] The displayed expression for D_KL(πθ||πref) appears nonstandard: the right-hand side resembles a first-order approximation or a reverse form rather than the usual KL divergence. Please correct the formula or explain the approximation being used.
  2. [Table 2 and Table 4] The captions use 'com / acc / rea' without defining the abbreviations; please define them in the captions or in the text near Sec. 3.1.
  3. [Sec. 3.1] The phrase 'export text annotations' should read 'expert text annotations'; this typo appears in the paragraph introducing the Q-Bench evaluation.
  4. [General] The paper states that the models are released but provides no code, data, model weights, or evaluation scripts, and the project page link is not included in the manuscript text. Please provide these artifacts or at least a detailed evaluation protocol so that the reasoning scores in Table 2 can be reproduced.
  5. [Appendix C (Fig. 5)] Figure 5(a) is described as 'Accuracy reward variation during training' but the text does not specify whether this is the mean R_acc over the rollout group or the smoothed group-level reward; please clarify the quantity plotted and its relationship to Eq. (5).

Circularity Check

2 steps flagged · score 4.0 of 10

Reasoning-description SOTA claim leans on a self-referential evaluation: the Gemini judge and the tested models both receive the ground-truth distortion tag, and 'Accuracy' is defined as agreement with that tag.

  1. self definitional [Sec. 3.1 (Quantitative Evaluation of Reasoning Capabilities); Appendix B Table 6; Appendix A Table 5]
    "Accuracy: Does the response correctly identify the dominant distortions and assign a severity level aligned with the ground-truth label? ... Task 2: Accuracy Evaluate whether the response correctly identifies key distortions—especially those relevant to the distortion tag—and whether the severity level is appropriate. ... I will give you a hint: the soft label for this image is [ distortion_tags ], which may suggest its potential distortion type and severity. You should carefully take this into account during the low-level attribute analysis."

    The reasoning-accuracy metric is defined as agreement with the ground-truth distortion tag, and that same tag is supplied to both the evaluated model and the judge. Appendix A gives the tag as a hint during distillation, and Sec. 4.2 says all evaluated models are guided using the same distillation templates, so the tag is effectively handed to the tested models. Appendix B then instructs Gemini-1.5-Pro to judge whether the response identifies distortions 'relevant to the distortion tag' and whether severity is 'appropriate' to that tag. A model can score high Accuracy by restating the provided hint rather than by visual reasoning.

  2. other [Sec. 3.1 (Quantitative Evaluation of Reasoning Capabilities)]
    "we enrich the semantic content of these labels by categorizing distortion severity into five levels: extreme, severe, noticeable, moderate, and slight, based on an equal partitioning of the MOS range [0, 1]."

    The severity reference used to grade Accuracy is constructed by the paper's own equal binning of the MOS range, not by human distortion-severity annotation. The same constructed severity label is inserted into the distillation prompt as a hint and then used as the correctness criterion in the evaluation prompt. Thus part of the 'aligned with the ground-truth label' check measures agreement with a label the paper itself generated and injected, so the claimed reasoning-accuracy gain is partially self-consistency with the paper's own label construction rather than agreement with an external standard.

full rationale

The score-regression half of the paper is self-contained and externally anchored: Table 1 reports PLCC/SRCC on human-rated benchmarks (KonIQ, SPAQ, LIVE-Wild, KADID, CSIQ, AGIQA), and Table 3 ablations show that adding long CoT and RL improves regression beyond SFT-only configurations. Those numbers do not reduce to the model's own inputs. The reasoning-description half, however, is partially circular. The Accuracy criterion is defined as agreement with the ground-truth distortion tag, and that tag is given both to the evaluated model as a hint and to the Gemini-1.5-Pro judge as the reference. Q-Ponder was distilled with this hint, so its high Accuracy score in Table 2 may largely reflect repeating the provided tag. The severity labels used as 'ground truth' are the paper's own MOS bins. The paper's Limitation explicitly concedes there is no human or second-MLLM verification, which weakens the claimed SOTA in reasoning descriptions. Overall score 4: the score-regression contribution is independent, but the unified claim of SOTA interpretability rests on a metric that is partly self-referential.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

The paper's central claims rest mainly on two unverified proxies: the teacher model for distillation quality and a single proprietary judge for reasoning quality, plus standard GRPO hyperparameters. No genuinely new physical or conceptual entities are introduced.

free parameters (8)
  • gamma (γ) = 0.8
    Retention ratio in heuristic filtering (Eq. 1); chosen by hand, and the paper shows filtering 20% improves distilled data correlation.
  • alpha (α) = 10
    Tolerance in accuracy reward (Eq. 5); smaller values penalize deviations more sharply, selected by the authors.
  • lambda_acc = 2
    Weight for accuracy reward in Eq. (7).
  • lambda_fmt = 1
    Weight for format reward in Eq. (7).
  • beta (β) = 0.04
    KL penalty coefficient in Eq. (2).
  • epsilon (ε) = 0.2
    Clipping coefficient in Eq. (2).
  • n = 8
    Number of sampled outputs per group in GRPO.
  • five-level severity scale = extreme, severe, noticeable, moderate, slight
    Equal partition of MOS [0,1] into five levels; a hand-defined discretization used for prompts and evaluation.
assumptions (4)
  • domain assumption Teacher model Qwen-2.5-VL-72B produces high-quality, factually grounded reasoning chains when prompted with the expert template.
    Cold-start distillation relies entirely on teacher outputs; filtering improves internal consistency but does not prove external validity.
  • domain assumption Gemini-1.5-Pro provides reliable completeness, accuracy, and reasonableness scores for reasoning descriptions.
    Authors state in Limitation that no human experts or other MLLMs verify the judge; this is the load-bearing evaluation assumption.
  • ad hoc to paper Supplying the ground-truth distortion tag as a hint during evaluation is a fair test of reasoning accuracy.
    This protocol is defined by the authors; it makes the accuracy metric partially a measure of hint-following.
  • standard math GRPO advantage normalization with max reward as baseline (Eq. 3) is a valid variant of GRPO.
    Taken from DeepSeek-R1 literature with a modification from mean to max; the paper does not analyze the effect of this change.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Q-Ponder: A Unified Training Pipeline for Reasoning-based Visual Quality Assessment." pith.science (2026). https://pith.science/paper/WIFEHNJ2

@misc{pith2026250605384,
  author       = {Pith},
  title        = {Pith review of: Q-Ponder: A Unified Training Pipeline for Reasoning-based Visual Quality Assessment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WIFEHNJ2}},
  note         = {Machine review of arXiv:2506.05384}
}
read the original abstract

Recent studies demonstrate that multimodal large language models (MLLMs) can proficiently evaluate visual quality through interpretable assessments. However, existing approaches typically treat quality scoring and reasoning descriptions as separate tasks with disjoint optimization objectives, leading to a trade-off: models adept at quality reasoning descriptions struggle with precise score regression, while score-focused models lack interpretability. This limitation hinders the full potential of MLLMs in visual quality assessment, where accuracy and interpretability should be mutually reinforcing. To address this, we propose a unified two-stage training framework comprising a cold-start stage and a reinforcement learning-based fine-tuning stage. Specifically, in the first stage, we distill high-quality data from a teacher model through expert-designed prompts, initializing reasoning capabilities via cross-entropy loss supervision. In the second stage, we introduce a novel reward with Group Relative Policy Optimization (GRPO) to jointly optimize scoring accuracy and reasoning consistency. We designate the models derived from these two stages as Q-Ponder-CI and Q-Ponder. Extensive experiments show that Q-Ponder achieves state-of-the-art (SOTA) performance on quality score regression benchmarks, delivering up to 6.5% higher SRCC on cross-domain datasets. Furthermore, Q-Ponder significantly outperforms description-based SOTA models, including its teacher model Qwen-2.5-VL-72B, particularly in description accuracy and reasonableness, demonstrating the generalization potential over diverse tasks.

Figures

Figures reproduced from arXiv: 2506.05384 by the authors.

Figure 1
Figure 1. Illustrations of the motivation. (a) A comprehensive, reasonable, and accurate "expert￾level" reasoning process helps the model regress precise quality scores, while maintaining a certain level of robustness on out-of-distribution data during training. (b) As the predicted score approaches the ground truth, the precision of the reasoning descriptions improves, indicating that it is possible to refine the model’s rea… view at source ↗
Figure 2
Figure 2. Illustration of the distillation process. (a) We design distillation prompts targeting both low-level (e.g., sharpness, color, artifacts) and high-level (e.g., composition, emotional expression) attributes, with enhanced distortion labeling (e.g., Severe_Blur) added to the dataset. (b) We filter generated reasoning chains based on the alignment between predicted and ground-truth scores. Completeness: Does the respon… view at source ↗
Figure 3
Figure 3. The policy model is first initialized via cold start to enhance reasoning. It then generates [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Illustration of main evaluation results. (a) Heatmap showing cross-dataset performance on quality score prediction (SRCC × PLCC). (b) Line plot comparing the completeness, accuracy, and reasonableness of reasoning descriptions across datasets for different models. unde…
Figure 5
Figure 5. Figure 5: Training dynamics of Q-Ponder during GRPO fine-tuning. (a) Accuracy reward variation [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Zoom-IQA: Image Quality Assessment with Reliable Region-Aware Reasoning

    cs.CV 2026-01 conditional novelty 7.0 of 10

    Zoom-IQA lets a vision-language model iteratively crop and zoom into image regions before giving a quality score, improving reasoning and restoration guidance over single-pass IQA models.

Reference graph

Works this paper leans on

143 extracted references · 57 canonical work pages · cited by 1 Pith paper

  1. [1]

    Flamingo: a visual language model for few-shot learning

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716–23736, 2022

    2022

  2. [2]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025

    arXiv 2025

  3. [3]

    Deep neural networks for no-reference and full-reference image quality assessment

    Sebastian Bosse, Dominique Maniry, Klaus-Robert Müller, Thomas Wiegand, and Wojciech Samek. Deep neural networks for no-reference and full-reference image quality assessment. IEEE Transactions on image processing, 27(1):206–219, 2017

    2017

  4. [4]

    Phocolens: Photorealistic and consistent reconstruction in lensless imaging

    Xin Cai, Zhiyuan You, Hailong Zhang, Jinwei Gu, Wentao Liu, and Tianfan Xue. Phocolens: Photorealistic and consistent reconstruction in lensless imaging. Advances in Neural Information Processing Systems, 37:12219–12242, 2024

    2024

  5. [5]

    A comprehensive survey of ai-generated content (aigc): A history of generative ai from gan to chatgpt

    Yihan Cao, Siyu Li, Yixin Liu, Zhiling Yan, Yutong Dai, Philip S Yu, and Lichao Sun. A comprehensive survey of ai-generated content (aigc): A history of generative ai from gan to chatgpt. arXiv preprint arXiv:2303.04226, 2023

    arXiv 2023

  6. [6]

    Q-ground: Image quality grounding with large multi-modality models

    Chaofeng Chen, Sensen Yang, Haoning Wu, Liang Liao, Zicheng Zhang, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Q-ground: Image quality grounding with large multi-modality models. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 486–495, 2024

    2024

  7. [7]

    Promptiqa: Boosting the performance and generalization for no-reference image quality assessment via prompts

    Zewen Chen, Haina Qin, Juan Wang, Chunfeng Yuan, Bing Li, Weiming Hu, and Liang Wang. Promptiqa: Boosting the performance and generalization for no-reference image quality assessment via prompts. In European Conference on Computer Vision, pages 247–264. Springer, 2024

    2024

  8. [8]

    Seagull: No-reference image quality assessment for regions of interest via vision-language instruction tuning

    Zewen Chen, Juan Wang, Wen Wang, Sunhan Xu, Hang Xiong, Yun Zeng, Jian Guo, Shuxun Wang, Chunfeng Yuan, Bing Li, et al. Seagull: No-reference image quality assessment for regions of interest via vision-language instruction tuning. arXiv preprint arXiv:2411.10161, 2024

    arXiv 2024

Show all 143 references
  1. [9]

    Study of subjective and objective naturalness assessment of ai-generated images.IEEE Transactions on Circuits and Systems for Video Technology, 2024

    Zijian Chen, Wei Sun, Haoning Wu, Zicheng Zhang, Jun Jia, Ru Huang, Xiongkuo Min, Guangtao Zhai, and Wenjun Zhang. Study of subjective and objective naturalness assessment of ai-generated images.IEEE Transactions on Circuits and Systems for Video Technology, 2024

    2024

  2. [10]

    Prompt-based test-time real image dehazing: a novel pipeline

    Zixuan Chen, Zewei He, Ziqian Lu, Xuecheng Sun, and Zhe-Ming Lu. Prompt-based test-time real image dehazing: a novel pipeline. In European Conference on Computer Vision, pages 432–449. Springer, 2024

    2024

  3. [11]

    Perceptual image quality assessment with transformers

    Manri Cheon, Sung-Jun Yoon, Byungyeon Kang, and Junwoo Lee. Perceptual image quality assessment with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 433–442, 2021

    2021

  4. [12]

    Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality

    Wei-Lin Chiang, Zhuohan Li, Ziqing Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality. See https://vicuna. lmsys. org (accessed 14 April 2023...

    2023

  5. [13]

    Deep rein- forcement learning from human preferences

    Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep rein- forcement learning from human preferences. Advances in neural information processing systems , 30, 2017

    2017

  6. [14]

    Perceptual quality assessment of smartphone photography

    Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3677–3686, 2020

    2020

  7. [15]

    Live in the wild image quality challenge database

    Deepti Ghadiyaram and Alan C Bovik. Live in the wild image quality challenge database. Online: http://live. ece. utexas. edu/research/ChallengeDB/index. html [Mar, 2017], 2015. 10

    2017

  8. [16]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

    2025 arXiv

  9. [17]

    Can we generate images with cot? let’s verify and reinforce image generation step by step

    Ziyu Guo, Renrui Zhang, Chengzhuo Tong, Zhizheng Zhao, Peng Gao, Hongsheng Li, and Pheng-Ann Heng. Can we generate images with cot? let’s verify and reinforce image generation step by step. arXiv preprint arXiv:2501.13926, 2025

    2025 arXiv

  10. [18]

    Image quality metrics: Psnr vs

    Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssim. In 2010 20th international conference on pattern recognition, pages 2366–2369. IEEE, 2010

    2010

  11. [19]

    Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment

    Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing, 29:4041–4056, 2020

    2020

  12. [20]

    Musiq: Multi-scale image quality transformer

    Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148–5157, 2021

    2021

  13. [21]

    Most apparent distortion: full-reference image quality assessment and the role of strategy

    Eric C Larson and Damon M Chandler. Most apparent distortion: full-reference image quality assessment and the role of strategy. Journal of electronic imaging, 19(1):011006–011006, 2010

    2010

  14. [22]

    Agiqa-3k: An open database for ai-generated image quality assessment

    Chunyi Li, Zicheng Zhang, Haoning Wu, Wei Sun, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. Agiqa-3k: An open database for ai-generated image quality assessment. IEEE Transactions on Circuits and Systems for Video Technology, 34(8):6833–6846, 2023

    2023

  15. [23]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023

    2023

  16. [24]

    Q-insight: Understanding image quality via visual reinforcement learning

    Weiqi Li, Xuanyu Zhang, Shijie Zhao, Yabin Zhang, Junlin Li, Li Zhang, and Jian Zhang. Q-insight: Understanding image quality via visual reinforcement learning. arXiv preprint arXiv:2503.22679, 2025

    2025 arXiv

  17. [25]

    Kadid-10k: A large-scale artificially distorted iqa database

    Hanhe Lin, Vlad Hosu, and Dietmar Saupe. Kadid-10k: A large-scale artificially distorted iqa database. In 2019 Eleventh International Conference on Quality of Multimedia Experience (QoMEX), pages 1–3. IEEE, 2019

    2019

  18. [26]

    Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023

    2023

  19. [27]

    Rankiqa: Learning from rankings for no- reference image quality assessment

    Xialei Liu, Joost Van De Weijer, and Andrew D Bagdanov. Rankiqa: Learning from rankings for no- reference image quality assessment. In Proceedings of the IEEE international conference on computer vision, pages 1040–1049, 2017

    2017

  20. [28]

    Cid: Iq–a new image quality database

    Xinwei Liu, Marius Pedersen, and Jon Yngve Hardeberg. Cid: Iq–a new image quality database. In Image and Signal Processing: 6th International Conference, ICISP 2014, Cherbourg, France, June 30–July 2,

    2014

  21. [29]

    Infigui-r1: Advancing multimodal gui agents from reactive actors to deliberative reasoners

    Yuhang Liu, Pengxiang Li, Congkai Xie, Xavier Hu, Xiaotian Han, Shengyu Zhang, Hongxia Yang, and Fei Wu. Infigui-r1: Advancing multimodal gui agents from reactive actors to deliberative reasoners. arXiv preprint arXiv:2504.14239, 2025

    2025 arXiv

  22. [30]

    Visual-rft: Visual reinforcement fine-tuning

    Ziyu Liu, Zeyi Sun, Yuhang Zang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. Visual-rft: Visual reinforcement fine-tuning. arXiv preprint arXiv:2503.01785, 2025

    2025 arXiv

  23. [31]

    Ui-r1: Enhancing action prediction of gui agents by reinforcement learning

    Zhengxi Lu, Yuxiang Chai, Yaxuan Guo, Xi Yin, Liang Liu, Hao Wang, Guanjing Xiong, and Hong- sheng Li. Ui-r1: Enhancing action prediction of gui agents by reinforcement learning. arXiv preprint arXiv:2503.21620, 2025

    2025 arXiv

  24. [32]

    Learning a no-reference quality metric for single-image super-resolution

    Chao Ma, Chih-Yuan Yang, Xiaokang Yang, and Ming-Hsuan Yang. Learning a no-reference quality metric for single-image super-resolution. Computer Vision and Image Understanding, 158:1–16, 2017

    2017

  25. [33]

    End-to-end blind image quality assessment using deep neural networks

    Kede Ma, Wentao Liu, Kai Zhang, Zhengfang Duanmu, Zhou Wang, and Wangmeng Zuo. End-to-end blind image quality assessment using deep neural networks. IEEE Transactions on Image Processing, 27(3):1202–1213, 2017

    2017

  26. [34]

    No-reference image quality assessment in the spatial domain

    Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on image processing, 21(12):4695–4708, 2012. 11

    2012

  27. [35]

    Making a ¡ °completely blind¡± image quality analyzer

    Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a ¡ °completely blind¡± image quality analyzer. IEEE Signal processing letters, 20(3):209–212, 2012

    2012

  28. [36]

    Human-level control through deep reinforcement learning

    V olodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. nature, 518(7540):529–533, 2015

    2015

  29. [37]

    A two-step framework for constructing blind image quality indices

    Anush Krishna Moorthy and Alan Conrad Bovik. A two-step framework for constructing blind image quality indices. IEEE Signal processing letters, 17(5):513–516, 2010

    2010

  30. [38]

    Blind image quality assessment: From natural scene statistics to perceptual quality

    Anush Krishna Moorthy and Alan Conrad Bovik. Blind image quality assessment: From natural scene statistics to perceptual quality. IEEE transactions on Image Processing, 20(12):3350–3364, 2011

    2011

  31. [39]

    Blind image quality assessment using a deep bilinear convolutional neural network

    A Deep Bilinear Convolutional Neural Network. Blind image quality assessment using a deep bilinear convolutional neural network. Deep Bilinear Convolutional Neural, 2022

    2022

  32. [40]

    Training language models to follow instructions with human feedback

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:2773...

    2022

  33. [41]

    Medvlm-r1: Incentivizing medical reasoning capability of vision-language models (vlms) via reinforcement learning

    Jiazhen Pan, Che Liu, Junde Wu, Fenglin Liu, Jiayuan Zhu, Hongwei Bran Li, Chen Chen, Cheng Ouyang, and Daniel Rueckert. Medvlm-r1: Incentivizing medical reasoning capability of vision-language models (vlms) via reinforcement learning. arXiv preprint arXiv:2502.19634, 2025

    2025 arXiv

  34. [42]

    Image database tid2013: Peculiarities, results and perspectives

    Nikolay Ponomarenko, Lina Jin, Oleg Ieremeiev, Vladimir Lukin, Karen Egiazarian, Jaakko Astola, Benoit V ozel, Kacem Chehdi, Marco Carli, Federica Battisti, et al. Image database tid2013: Peculiarities, results and perspectives. Signal processing: Image communication, 30:57–77, 2015

    2015

  35. [43]

    Tid2008-a database for evaluation of full-reference visual quality assessment metrics

    Nikolay Ponomarenko, Vladimir Lukin, Alexander Zelensky, Karen Egiazarian, Marco Carli, and Federica Battisti. Tid2008-a database for evaluation of full-reference visual quality assessment metrics. Advances of modern radioelectronics, 10(4):30–45, 2009

    2009

  36. [44]

    Pieapp: Perceptual image-error assessment through pairwise preference

    Ekta Prashnani, Hong Cai, Yasamin Mostofi, and Pradeep Sen. Pieapp: Perceptual image-error assessment through pairwise preference. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1808–1817, 2018

    2018

  37. [45]

    Blind image quality assessment: A natural scene statistics approach in the dct domain

    Michele A Saad, Alan C Bovik, and Christophe Charrier. Blind image quality assessment: A natural scene statistics approach in the dct domain. IEEE transactions on Image Processing, 21(8):3339–3352, 2012

    2012

  38. [46]

    Proximal policy optimization algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  39. [47]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024

    2024 arXiv

  40. [48]

    Image information and visual quality

    Hamid R Sheikh and Alan C Bovik. Image information and visual quality. IEEE Transactions on image processing, 15(2):430–444, 2006

    2006

  41. [49]

    A statistical evaluation of recent full reference image quality assessment algorithms

    Hamid R Sheikh, Muhammad F Sabir, and Alan C Bovik. A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Transactions on image processing, 15(11):3440–3451, 2006

    2006

  42. [50]

    Vlm-r1: A stable and generalizable r1-style large vision-language model

    Haozhan Shen, Peng Liu, Jingcheng Li, Chunxin Fang, Yibo Ma, Jiajia Liao, Qiaoli Shen, Zilun Zhang, Kangjia Zhao, Qianqian Zhang, et al. Vlm-r1: A stable and generalizable r1-style large vision-language model. arXiv preprint arXiv:2504.07615, 2025

    2025 arXiv

  43. [51]

    Hybridflow: A flexible and efficient rlhf framework

    Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. Hybridflow: A flexible and efficient rlhf framework. arXiv preprint arXiv:2409.19256, 2024

    2024 arXiv

  44. [52]

    Mastering the game of go with deep neural networks and tree search

    David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search. nature, 529(7587):484–489, 2016

    2016

  45. [53]

    Blindly assess image quality in the wild guided by a self-adaptive hyper network

    Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3667–3676, 2020. 12

    2020

  46. [54]

    Graphiqa: Learning distortion graph representations for blind image quality assessment

    Simeng Sun, Tao Yu, Jiahua Xu, Wei Zhou, and Zhibo Chen. Graphiqa: Learning distortion graph representations for blind image quality assessment. IEEE Transactions on Multimedia, 25:2912–2925, 2022

    2022

  47. [55]

    Nima: Neural image assessment

    Hossein Talebi and Peyman Milanfar. Nima: Neural image assessment. IEEE transactions on image processing, 27(8):3998–4011, 2018

    2018

  48. [56]

    Learning a blind measure of perceptual image quality

    Huixuan Tang, Neel Joshi, and Ashish Kapoor. Learning a blind measure of perceptual image quality. In CVPR 2011, pages 305–312. IEEE, 2011

    2011

  49. [57]

    Internlm: A multilingual language model with progressively enhanced capabilities, 2023

    InternLM Team. Internlm: A multilingual language model with progressively enhanced capabilities, 2023

    2023

  50. [58]

    Gpt-4v (ision) system card

    OpenAI Team. Gpt-4v (ision) system card. 2023. URL https://api. semanticscholar. org/CorpusID , 263218031, 2024

    2023

  51. [59]

    Exploring clip for assessing the look and feel of images

    Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI conference on artificial intelligence, volume 37, pages 2555–2563, 2023

    2023

  52. [60]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600–612, 2004

    2004

  53. [61]

    Mp-gui: Modality perception with mllms for gui understanding

    Ziwei Wang, Weizhi Chen, Leyang Yang, Sheng Zhou, Shengchu Zhao, Hanbei Zhan, Jiongchao Jin, Liangcheng Li, Zirui Shao, and Jiajun Bu. Mp-gui: Modality perception with mllms for gui understanding. arXiv preprint arXiv:2503.14021, 2025

    2025 arXiv

  54. [62]

    Q-bench: A benchmark for general-purpose foundation models on low-level vision

    Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Chunyi Li, Wenxiu Sun, Qiong Yan, Guangtao Zhai, et al. Q-bench: A benchmark for general-purpose foundation models on low-level vision. arXiv preprint arXiv:2309.14181, 2023

    2023 arXiv

  55. [63]

    Q-instruct: Improving low-level visual abilities for multi-modality foundation models

    Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Kaixin Xu, Chunyi Li, Jingwen Hou, Guangtao Zhai, et al. Q-instruct: Improving low-level visual abilities for multi-modality foundation models. In Proceedings of the IEEE/CVF conference on computer v...

    2024

  56. [64]

    Towards open-ended visual quality comparison

    Haoning Wu, Hanwei Zhu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Annan Wang, Wenxiu Sun, Qiong Yan, et al. Towards open-ended visual quality comparison. In European Conference on Computer Vision, pages 360–377. Springer, 2024

    2024

  57. [65]

    A comprehensive study of multimodal large language models for image quality assessment

    Tianhe Wu, Kede Ma, Jie Liang, Yujiu Yang, and Lei Zhang. A comprehensive study of multimodal large language models for image quality assessment. In European Conference on Computer Vision, pages 143–160. Springer, 2024

    2024

  58. [66]

    Maniqa: Multi-dimension attention network for no-reference image quality assessment

    Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniqa: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, page...

    2022

  59. [67]

    mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration

    Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. In Proceedings of the ieee/cvf conference on computer vision and pattern recognition, p...

    2024

  60. [68]

    Teaching large language models to regress accurate image quality scores using score distribution

    Zhiyuan You, Xin Cai, Jinjin Gu, Tianfan Xue, and Chao Dong. Teaching large language models to regress accurate image quality scores using score distribution. arXiv preprint arXiv:2501.11561, 2025

    2025

  61. [69]

    Descriptive image quality assessment in the wild

    Zhiyuan You, Jinjin Gu, Zheyuan Li, Xin Cai, Kaiwen Zhu, Chao Dong, and Tianfan Xue. Descriptive image quality assessment in the wild. arXiv preprint arXiv:2405.18842, 2024

    2024

  62. [70]

    Depicting beyond scores: Advancing image quality assessment through multi-modal language models

    Zhiyuan You, Zheyuan Li, Jinjin Gu, Zhenfei Yin, Tianfan Xue, and Chao Dong. Depicting beyond scores: Advancing image quality assessment through multi-modal language models. In European Conference on Computer Vision, pages 259–276. Springer, 2024

    2024

  63. [71]

    Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback

    Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, et al. Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. In Proceedings of the IEEE/CVF Conference on Co...

    2024

  64. [72]

    Rlaif-v: Aligning mllms through open-source ai feedback for super gpt-4v trustworthiness

    Tianyu Yu, Haoye Zhang, Yuan Yao, Yunkai Dang, Da Chen, Xiaoman Lu, Ganqu Cui, Taiwen He, Zhiyuan Liu, Tat-Seng Chua, et al. Rlaif-v: Aligning mllms through open-source ai feedback for super gpt-4v trustworthiness. arXiv preprint arXiv:2405.17220, 2024

    2024

  65. [73]

    Sf-iqa: Quality and similarity integration for ai generated image quality assessment

    Zihao Yu, Fengbin Guan, Yiting Lu, Xin Li, and Zhibo Chen. Sf-iqa: Quality and similarity integration for ai generated image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6692–6701, 2024

    2024

  66. [74]

    Vision-language models for vision tasks: A survey

    Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  67. [75]

    Fsim: A feature similarity index for image quality assessment

    Lin Zhang, Lei Zhang, Xuanqin Mou, and David Zhang. Fsim: A feature similarity index for image quality assessment. IEEE transactions on Image Processing, 20(8):2378–2386, 2011

    2011

  68. [76]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586–595, 2018

    2018

  69. [77]

    Continual learning for blind image quality assessment

    Weixia Zhang, Dingquan Li, Chao Ma, Guangtao Zhai, Xiaokang Yang, and Kede Ma. Continual learning for blind image quality assessment. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):2864–2878, 2022

    2022

  70. [78]

    Blind image quality assessment via vision-language correspondence: A multitask learning perspective

    Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14071–14081, 2023

    2023

  71. [79]

    Gpt-4v (ision) as a generalist evaluator for vision-language tasks

    Xinlu Zhang, Yujie Lu, Weizhi Wang, An Yan, Jun Yan, Lianke Qin, Heng Wang, Xifeng Yan, William Yang Wang, and Linda Ruth Petzold. Gpt-4v (ision) as a generalist evaluator for vision-language tasks. arXiv preprint arXiv:2311.01361, 2023

    2023 arXiv

  72. [80]

    Mm-rlhf: The next step forward in multimodal llm alignment

    Yi-Fan Zhang, Tao Yu, Haochen Tian, Chaoyou Fu, Peiyan Li, Jianshu Zeng, Wulin Xie, Yang Shi, Huanyu Zhang, Junkang Wu, et al. Mm-rlhf: The next step forward in multimodal llm alignment. arXiv preprint arXiv:2502.10391, 2025

    2025 arXiv

  73. [81]

    Q-bench-video: Benchmarking the video quality understanding of lmms

    Zicheng Zhang, Ziheng Jia, Haoning Wu, Chunyi Li, Zijian Chen, Yingjie Zhou, Wei Sun, Xiaohong Liu, Xiongkuo Min, Weisi Lin, et al. Q-bench-video: Benchmarking the video quality understanding of lmms. arXiv preprint arXiv:2409.20063, 2024

    2024 arXiv

  74. [82]

    Quality assess- ment in the era of large models: A survey

    Zicheng Zhang, Yingjie Zhou, Chunyi Li, Baixuan Zhao, Xiaohong Liu, and Guangtao Zhai. Quality assess- ment in the era of large models: A survey. ACM Transactions on Multimedia Computing, Communications and Applications, 2024

    2024

  75. [83]

    Learning conditional knowledge distillation for degraded-reference image quality assessment

    Heliang Zheng, Huan Yang, Jianlong Fu, Zheng-Jun Zha, and Jiebo Luo. Learning conditional knowledge distillation for degraded-reference image quality assessment. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 10242–10251, 2021

    2021

  76. [84]

    Causal-iqa: Towards the generaliza- tion of image quality assessment based on causal inference

    Yan Zhong, Xingyu Wu, Li Zhang, Chenxi Yang, and Tingting Jiang. Causal-iqa: Towards the generaliza- tion of image quality assessment based on causal inference. In Forty-first International Conference on Machine Learning, 2024

    2024

  77. [85]

    Quality assessment of image super-resolution: Balancing deterministic and statistical fidelity

    Wei Zhou and Zhou Wang. Quality assessment of image super-resolution: Balancing deterministic and statistical fidelity. In Proceedings of the 30th ACM international conference on multimedia, pages 934–942, 2022

    2022

  78. [86]

    Metaiqa: Deep meta-learning for no-reference image quality assessment

    Hancheng Zhu, Leida Li, Jinjian Wu, Weisheng Dong, and Guangming Shi. Metaiqa: Deep meta-learning for no-reference image quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14143–14152, 2020

    2020

  79. [87]

    Adaptive image quality assessment via teaching large multimodal model to compare

    Hanwei Zhu, Haoning Wu, Yixuan Li, Zicheng Zhang, Baoliang Chen, Lingyu Zhu, Yuming Fang, Guangtao Zhai, Weisi Lin, and Shiqi Wang. Adaptive image quality assessment via teaching large multimodal model to compare. arXiv preprint arXiv:2405.19298, 2024

    2024 arXiv

  80. [88]

    An intelligent agentic system for complex image restoration problems

    Kaiwen Zhu, Jinjin Gu, Zhiyuan You, Yu Qiao, and Chao Dong. An intelligent agentic system for complex image restoration problems. arXiv preprint arXiv:2410.17809, 2024. 14 A Prompt Templates for Reasoning Distillation Prompt Structure Overview. The table below summarizes the s...

    2024 arXiv

  81. [90]

    Brightness and Exposure

  82. [91]

    Sharpness and Detail Preservation

  83. [92]

    compression noise)

    Noise Level (digital vs. compression noise)

  84. [93]

    Color Performance (WB, saturation, transitions)

  85. [94]

    Distortion and Artifacts (e.g., lens, moiré)

  86. [95]

    Special rule: Artistic style may conditionally override low-level flaws, but must not affect final score

    Visual Interference (e.g., glare, weather, clutter) High-Level Attribute Analysis Note: These are supportive attributes, subordinate to low-level analysis. Special rule: Artistic style may conditionally override low-level flaws, but must not affect final score

  87. [96]

    Subject Clarity (focus, separation)

  88. [97]

    Composition and Layout (symmetry, rule of thirds, balance)

  89. [98]

    Emotional Expression and Storytelling

  90. [99]

    Artistic Style and Creativity

  91. [100]

    Environment and Background Integration

  92. [101]

    landscape) Comprehensive Eval- uation and Output Format Summarize the analysis and provide concrete improvement suggestions

    Context Adaptability (e.g., portrait vs. landscape) Comprehensive Eval- uation and Output Format Summarize the analysis and provide concrete improvement suggestions. Clearly explain the reasoning basis and logic. Structure the output into three numbered sections: (1) Low-level...

    arXiv

  93. [102]

    The details of the tiger’s fur and facial features are not very sharp

    Clarity: The image appears to be somewhat blurry, which affects the overall clarity. The details of the tiger’s fur and facial features are not very sharp

  94. [103]

    This makes it difficult to discern fine details

    Focus: The focus seems to be off, as the image lacks sharpness in both the foreground and background. This makes it difficult to discern fine details

  95. [104]

    Lighting: The lighting is adequate, but there might be some overexposure in certain areas, particularly around the tiger’s face, which could be due to the bright background

  96. [105]

    Color: The colors appear natural, but the image has a high level of noise or graininess, which can be distracting and reduce the overall quality

  97. [106]

    Composition: The composition is centered, which is acceptable, but the lack of sharpness detracts from the visual impact

  98. [107]

    Considering these factors, the image quality is not high

    Noise/Gain: There is noticeable noise or gain in the image, which is evident in the grainy texture, especially in the darker areas. Considering these factors, the image quality is not high. It suffers from blurriness, poor focus, and excessive noise, which significantly degrad...

  99. [108]

    There is a significant portion of the image that seems overexposed, particularly in the lighter areas like the tiger’s fur and background, resulting in a loss of detail

    In the analysis of low-level attributes, The image’s brightness and exposure appear uneven. There is a significant portion of the image that seems overexposed, particularly in the lighter areas like the tiger’s fur and background, resulting in a loss of detail. Conversely, som...

  100. [109]

    In the analysis of high-level attributes, the subject clarity is compromised as the tiger, although central, lacks sharpness and detail due to the noise and overexposure. The composition and layout do not effectively utilize established principles like the rule of thirds or le...

  101. [110]

    The severe denoising distortion, overexposure, and loss of detail significantly degrade the image quality

    Comprehensively evaluating the image, it suffers heavily from technical faults that overshadow any potential aesthetic qualities. The severe denoising distortion, overexposure, and loss of detail significantly degrade the image quality. Improvements should focus on adjusting e...

  102. [111]

    There is no noticeable blurriness or distortion that would affect the overall clarity

    Clarity: The image appears to be clear, with distinct lines and shapes visible in the architectural structure. There is no noticeable blurriness or distortion that would affect the overall clarity

  103. [112]

    The perspective used emphasizes the depth and curvature of the building, making it engaging

    Composition: The composition is well-balanced, with the circular structure creating a strong visual interest. The perspective used emphasizes the depth and curvature of the building, making it engaging

  104. [113]

    There are no harsh shadows or overexposed areas, which suggests good exposure management

    Lighting: The lighting is even, with natural light illuminating the facade of the building. There are no harsh shadows or overexposed areas, which suggests good exposure management

  105. [114]

    This adds to the aesthetic appeal of the image

    Color: The colors are natural and well-saturated, with a warm tone dominating due to the material of the building. This adds to the aesthetic appeal of the image

  106. [115]

    Based on these observations, the image quality is high

    Detail: The details of the building’s facade are sharp and well-defined, allowing for a clear view of the architectural features. Based on these observations, the image quality is high. Here is the assigned score: Score: 95.00 Q-Ponder-CI

  107. [116]

    However, some areas near the light source are slightly overexposed, leading to minor detail loss

    The image’s brightness and exposure seem appropriate for the scene, capturing the interplay of light and shadow across the building’s facade. However, some areas near the light source are slightly overexposed, leading to minor detail loss. Contrast is well-managed, providing a...

  108. [117]

    The composition employs leading lines created by the building’s curved facade, guiding the viewer’s eye towards the center where the light is concentrated

    The subject, which is the architectural structure, is clearly defined against the sky. The composition employs leading lines created by the building’s curved facade, guiding the viewer’s eye towards the center where the light is concentrated. This contributes to a balanced and...

  109. [118]

    Improvements could involve fine-tuning exposure settings to prevent highlight clipping and using HDR techniques to recover shadow details

    Overall, the image displays good quality with only minor issues in overexposure and shadow detail loss. Improvements could involve fine-tuning exposure settings to prevent highlight clipping and using HDR techniques to recover shadow details. Enhancing the contrast subtly can ...

  110. [119]

    This can make it difficult to discern fine details

    Clarity and Focus: The image appears somewhat blurry, which suggests that it may not be in sharp focus. This can make it difficult to discern fine details

  111. [120]

    This can wash out details and create glare

    Lighting: The lighting seems natural but is somewhat overexposed in certain areas, particularly around the center where the light is brightest. This can wash out details and create glare

  112. [121]

    However, the saturation could be adjusted for better balance

    Color Saturation: The colors appear quite vivid, especially the greens, which might be due to the natural environment or post-processing effects. However, the saturation could be adjusted for better balance

  113. [122]

    However, the overexposure in the center might distract from the overall composition

    Composition: The composition is interesting, with a natural frame created by the trees and a path leading into the scene. However, the overexposure in the center might distract from the overall composition

  114. [123]

    Noise and Grain: There appears to be some noise or graininess, particularly noticeable in the darker areas, which can reduce the perceived quality of the image

  115. [124]

    Detail Retention: Due to the blurriness and overexposure, there is a loss of detail in the image, making it harder to appreciate finer elements of the scene. Considering these points, here is a step-by-step evaluation: Clarity and Focus: 30/100 (low clarity) Lighting: 40/100 (...

  116. [125]

    The brighter areas near the top where sunlight filters through the trees seem slightly overexposed, causing some loss of detail

    The image’s brightness and exposure show a mix of challenges. The brighter areas near the top where sunlight filters through the trees seem slightly overexposed, causing some loss of detail. Conversely, the lower parts of the image are underexposed, resulting in a lack of visi...

  117. [126]

    The composition uses leading lines formed by the tree trunks and branches, subtly guiding the viewer’s gaze through the image

    Subject clarity is impacted by the lack of sharpness and detail, making it difficult to distinguish specific elements within the scene. The composition uses leading lines formed by the tree trunks and branches, subtly guiding the viewer’s gaze through the image. However, the l...

  118. [127]

    Reducing exposure in the brighter spots and increasing fill light in the darker areas would help achieve a more balanced exposure

    Considering all aspects, the image has moderate quality with several areas needing improvement. Reducing exposure in the brighter spots and increasing fill light in the darker areas would help achieve a more balanced exposure. Adjusting contrast and sharpening the image could ...

  119. [128]

    The details are sharp, especially in the foreground where the pencils are in focus

    Clarity: The image appears to be clear, with distinct edges and shapes visible in the pencils. The details are sharp, especially in the foreground where the pencils are in focus

  120. [129]

    This creates a pleasing depth of field that draws attention to the main subject

    Focus: The image is focused on the pencils in the foreground, while the background is intentionally blurred (bokeh effect). This creates a pleasing depth of field that draws attention to the main subject

  121. [130]

    The pastel shades of the pencils are visually appealing and add to the aesthetic quality

    Color: The colors are vibrant and well-saturated, contributing to the overall appeal of the image. The pastel shades of the pencils are visually appealing and add to the aesthetic quality

  122. [131]

    The arrangement of the pencils in the foreground and the blurred background creates a sense of depth and balance

    Composition: The composition is simple yet effective. The arrangement of the pencils in the foreground and the blurred background creates a sense of depth and balance. The angle at which the pencils are placed adds visual interest

  123. [132]

    Noise: There is no noticeable noise or graininess in the image, indicating good image quality

  124. [133]

    The use of bokeh and the vibrant colors enhance the visual impact

    Overall Impression: The image is aesthetically pleasing and effectively conveys its subject matter. The use of bokeh and the vibrant colors enhance the visual impact. Based on these observations, I would assign the image a quality score of 95.00. 24 Q-Ponder-CI 1.In the low-le...

  125. [134]

    The composition and layout appear to aim for a scattered, casual arrangement of the pencils, but the lack of sharpness hinders the effectiveness of this intended randomness

    Moving to the high-level attribute analysis, the subject clarity suffers significantly due to the blurriness, making it challenging to distinguish the main subject from the background. The composition and layout appear to aim for a scattered, casual arrangement of the pencils,...

  126. [135]

    While the color palette offers a pleasant pastel aesthetic, the lack of sharpness overshadows this positive aspect, leading to an overall diminished image quality

    Comprehensively evaluating the image quality, the severe global blur has a substantial negative impact across both low-level and high-level attributes. While the color palette offers a pleasant pastel aesthetic, the lack of sharpness overshadows this positive aspect, leading t...

  127. [136]

    Resolution: The image appears to have a high resolution, as there are no visible pixelation or blurriness that would suggest lower quality

  128. [137]

    The background is also well-defined without any noticeable artifacts

    Clarity: The image is very clear, with sharp details in the subject’s hair, facial features, and clothing. The background is also well-defined without any noticeable artifacts

  129. [138]

    There are no harsh shadows or overexposed areas

    Lighting: The lighting is soft and even, which enhances the natural look of the subject’s skin tone and hair texture. There are no harsh shadows or overexposed areas

  130. [139]

    The skin tones appear natural and healthy

    Color Balance: The colors are well-balanced, with a pleasant pastel palette in the background that complements the subject’s attire. The skin tones appear natural and healthy

  131. [140]

    This level of detail suggests a high-quality capture

    Detail: The image captures fine details, such as the individual strands of hair and the texture of the fabric. This level of detail suggests a high-quality capture

  132. [141]

    Based on these observations, the image quality is excellent

    Composition: The composition is well-balanced, with the subject positioned centrally and the background providing a nice contrast without distracting elements. Based on these observations, the image quality is excellent. Therefore, I would assign it a score of 98.00 out of 100...

  133. [142]

    The subject’s details and contours are visible, contributing to a clear understanding of what the image represents

    Subject clarity is good, with the main subject being distinct from the background and adequately focused. The subject’s details and contours are visible, contributing to a clear understanding of what the image represents. Composition and layout follow basic aesthetic principle...

  134. [143]

    While there is room for improvement in terms of sharpness and detail, the current state is acceptable given the moderate level of exposure and color accuracy

    Overall, the image displays a good balance of technical and aesthetic attributes. While there is room for improvement in terms of sharpness and detail, the current state is acceptable given the moderate level of exposure and color accuracy. To enhance the image further, increa...

  135. [2014]

    Springer, 2014

    Proceedings 6, pages 193–202. Springer, 2014

    2014

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.