Pith. sign in

REVIEW 3 major objections 5 minor 62 references

AU-IQA: A Benchmark Dataset for Perceptual Quality Assessment of AI-Enhanced User-Generated Content

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read User-photo quality models judge AI-enhanced images best

desk verdict A useful new IQA benchmark for AI-enhanced images, but the headline claim that AI-UGC resembles UGC rests on a synthetic pipeline that isn't validated against real AI-UGC. read the letter →

arxiv 2508.05016 v2 pith:2FWJAKQE submitted 2025-08-07 cs.CV eess.IV

Shushi Wang , Chunyi Li , Zicheng Zhang , Han Zhou , Wei Dong , Jun Chen , Guangtao Zhai , Xiaohong Liu This is my paper
classification cs.CVeess.IV
keywords AI-EnhancedUser-GeneratedContentPerceptualQualityAssessmentbenchmarkdatasetmeanopinionscoresuper-resolutionlow-lightenhancementdenoisinglargemultimodalmodels
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

The paper argues that AI-enhanced user-generated content (AI-UGC) — photos that platforms or users have upscaled, denoised, or brightened with deep learning — forms a quality-assessment category of its own, distinct from both ordinary UGC and fully AI-generated images. To make this category testable, the authors build AU-IQA, a benchmark of 4,800 images with human quality scores: 400 high-quality photos from an existing dataset are artificially degraded by downsampling, darkening, or Gaussian noise, then restored by four advanced enhancement models per degradation type. On this benchmark they evaluate a range of existing quality models and find that models trained on traditional user-generated content outperform models trained on AI-generated content, and outperform AIGC-specialized models by a wide margin. They conclude that AI-UGC perceptually resembles UGC more than AIGC, while noting that model performance varies widely across enhancement types and becomes more stable on larger test sets.

What carries the argument

The load-bearing object is the AU-IQA dataset itself, produced by a synthetic degradation–restoration pipeline. Starting from 400 high-quality UGC photos from KonIQ-10k, three controlled degradations (downsampling, brightness reduction, Gaussian noise) are applied, and each degraded image is passed through four enhancement models corresponding to its degradation type, producing 4,800 AI-UGC images with MOS annotations. The benchmarking protocol — comparing PLCC and SRCC of UGC-trained, AIGC-trained, and large-multimodal-model predictors on per-type, random-subset, and full-dataset splits — is what carries the conclusion that UGC-trained models are the best current estimators of AI-UGC qualit

What would settle it

Collect a set of real AI-enhanced UGC images from platforms (e.g., phone low-light shots passed through denoising or enhancement apps, upscaled legacy photos) with MOS labels; compute feature statistics (e.g., from a pretrained visual embedding) and compare AU-IQA image features and model rankings to this real set. A large distributional gap or a reordering of model performance would falsify the claim that the synthetic pipeline captures AI-UGC quality.

Watch

Extended reading notes

Core claim

The central claim is that AU-IQA is a valid public benchmark for AI-UGC perceptual quality, and that the first benchmark results establish a usable baseline: UGC-focused IQA models generalize best to AI-enhanced content. The dataset is constructed by taking 400 high-quality KonIQ-10k images, applying three controlled degradations — low resolution, low light, and Gaussian noise — and restoring each degraded image with four enhancement models per type (DiffBIR, OSEDiff, PASD, SUPIR for super-resolution; GLARE, LightenDiffusion, NeRCo, QuadPrior for low-light; DiffBIR, MaskedDenoising, PASD, SUPIR for denoising), yielding 4,800 images labeled with mean opinion scores from five trained annotator

Load-bearing premise

The benchmark's validity rests on the assumption that images created by downsampling, darkening, or adding Gaussian noise to pristine photos and then restoring them with current enhancement models are perceptually representative of the AI-enhanced user content that actually appears on platforms.

Editorial extensions

If this is right

  • AI-enhanced user content can, to a first approximation, be scored by existing blind IQA models trained on ordinary UGC, so practitioners do not need to wait for a dedicated model before adding perceptual monitoring to enhancement pipelines.
  • AIGC-specialized quality models are a weak starting point for AI-UGC and should be retrained or adapted before use in this setting.
  • Low-light enhanced images are the hardest subtype for current models, pointing to a concrete target for new training data or model design.
  • Model rankings shift between per-type and full-dataset evaluation, so future benchmarks should report both small-scale and large-scale results to expose instability.
  • The dataset provides a controlled testbed for developing AI-UGC-specific quality models and for fine-tuning multimodal models.

Reading between the lines

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

  • Because the seed images are pristine KonIQ-10k photos, the conclusion that AI-UGC perceptually resembles UGC may be partly a consequence of construction; real AI-UGC from platforms carries authentic degradations that the three synthetic manipulations do not reproduce. A direct feature-distribution comparison between AU-IQA and scraped real AI-UGC would test this.
  • If the synthetic pipeline under-represents real-world degradations, the benchmark could still be used as a controlled stress test, but its absolute MOS values and model rankings should not be taken as platform-accurate without validation.
  • The observed improvement on larger test sets suggests models are partially fitting linear trends rather than true perceptual quality; reporting confidence intervals or per-image residuals would make the instability measurable and comparable.
  • A natural extension is to train a lightweight AI-UGC-specific regressor on AU-IQA and check whether it beats UGC models on genuinely enhanced content; that would directly test whether the category deserves its own model class.
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

3 major / 5 minor

Summary. The paper defines AI-UGC as user-generated content processed by AI enhancement models and introduces AU-IQA, a 4,800-image perceptual quality benchmark. The dataset is built by taking 400 high-quality KonIQ-10k photos, applying three controlled degradations (downsampling, brightness reduction, Gaussian noise), and restoring them with nine enhancement models spanning super-resolution, low-light enhancement, and denoising. Each image is annotated by five trained annotators and assigned a MOS. The authors evaluate a range of no-reference IQA models, LMM-based scorers, and LMMs fine-tuned on UGC, AIGC, or hybrid data, reporting PLCC and SRCC per enhancement model and per enhancement type. The central empirical conclusion is that UGC-trained quality models outperform AIGC-trained and hybrid models on AU-IQA, suggesting that AI-UGC perceptually resembles UGC more than AIGC, while also noting instability across enhancement types and test-set sizes.

Significance. If the conclusions hold, AU-IQA fills a real gap: there is currently no dedicated public benchmark for AI-enhanced user-generated content, and the paper provides a reasonably large MOS-labeled dataset together with a systematic comparison of existing IQA approaches. The dataset link is provided, which is a concrete contribution to the community. The manuscript also makes a useful practical claim that UGC-trained models are a better starting point than AIGC-trained models for this type of content. However, the strength of this claim depends on the ecological validity of the synthetic degradation pipeline and on the statistical reliability of the reported correlations; both need additional support before the qualitative ordering can be treated as a robust finding.

major comments (3)
  1. [Section 3.2] Construct validity of the AI-UGC stimulus set is the load-bearing premise. All 4,800 images are synthesized from 400 high-quality KonIQ-10k photos using only three controlled manipulations (downsampling, brightness reduction, Gaussian noise) followed by restoration. The paper explicitly adopts this as a compromise for 'controlled analysis,' but the central conclusion—that AI-UGC quality assessment is 'more similar to UGC' and UGC-trained models perform best—depends on these synthetic images standing in for real AI-UGC. Real UGC also contains compression, motion blur, sensor noise, uneven illumination, and mixed distortions, and often starts from lower-quality originals. Because the UGC-finetuned baselines include models trained on KADID-10K, whose synthetic distortions overlap with the degradation families used here, the observed advantage may reflect train/test distortion overlap rather
  2. [Section 3.2 (Annotation)] No inter-rater agreement is reported. Each image is rated by five annotators, with 'cross-checked for consistency' and relabeling in cases of significant discrepancies, but no ICC, Krippendorff's alpha, or per-image score variance is given. For a dataset whose benchmark value is the MOS, unreported annotator reliability is a load-bearing gap: if agreement is low, all downstream correlations are diluted. Please report agreement statistics, describe the relabeling rule, and provide the distribution of raw individual scores (or at least per-image confidence intervals).
  3. [Section 4.2, Tables 1-2] The model-ranking claims are made without statistical inference. Many correlations are low (e.g., ARNIQA PLCC = -0.074 on DiffBIR SR and -0.260 on denoising in Table 1; Q-Eval-Score correlations are around 0.25 throughout Table 2), and the differences between the UGC, AIGC, and hybrid groups are often small. The 'Random' column in Table 2 appears to be a single random draw of 1,600 images; no repetitions, seeds, confidence intervals, or significance tests are reported. The statement that models perform better on the larger 'All' set 'may result from the increased data volume making it easier for models to fit underlying linear relationships' is speculative without error bars. Please add bootstrap CIs or significance tests, and report repeated random splits for the 'Random' condition.
minor comments (5)
  1. [Section 3.2] Grammar: 'The annotations are required cross-checked' should be 'The annotations were cross-checked.' Also, please state the exact criterion for selecting 'high-quality samples' from KonIQ-10k (e.g., the MOS threshold or percentile).
  2. [Table 1 / Figure 4] The caption and axis text appear garbled in the version I reviewed (placeholder characters in Table 1's caption and unreadable axis labels in Figure 4). Please ensure the camera-ready formatting is legible.
  3. [Section 4.1] The grouping description is confusing: the text says that only MA-AGIQA is selected as an AIGC quality assessment model because most AIGC models use SBS evaluation, but then several LMMs fine-tuned on AGIQA-3K are also treated as AIGC models. Please clarify the grouping criteria.
  4. [Section 2.1] The text cites 'IR-SDE [26]' for diffusion-based restoration, but reference [26] appears to be 'Refusion.' Please verify the citation-reference match.
  5. [Section 4.2] The phrase 'the quality assessment of low-light enhanced images appears to be more consistent across two other enhancement types' is unclear; presumably 'more consistent than the other two enhancement types' is intended. Please rephrase.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor self-definitional circularity: AI-UGC is constructed from UGC seeds, so the finding that it resembles UGC is partly inherited from the construction.

  1. self definitional [Section 3.2 (Data Collection) and Section 4.2 (Performance Discussion)]
    "To enable systematic evaluation and controlled analysis, we adopt a synthetic degradation strategy based on high-quality UGC images. Specifically, we select 400 high-quality samples from the KonIQ-10k dataset... Three common degradation types—low resolution, low-light conditions, and noise—are then applied... This suggests that the features of AI-UGC images are more similar to those of traditional UGC images than to AIGC images."

    The dataset defines AI-UGC operationally as enhanced UGC: the seed images are taken from KonIQ-10k (a UGC dataset) and then degraded and restored by AI models. The subsequent conclusion that 'the features of AI-UGC images are more similar to those of traditional UGC images' is therefore not an independent empirical discovery but a direct consequence of the construction—the enhanced images inherit the content statistics of the UGC seeds. Additionally, the UGC baselines were finetuned on KADID-10K, whose distortion types (including Gaussian noise and brightness changes) overlap with the synthetic degradations used to create AU-IQA, so the model-ranking advantage of UGC-trained models may partly reflect train/test distortion-family overlap rather than a general perceptual property of real-wor

full rationale

The paper is primarily a dataset and benchmarking contribution. The construction of AU-IQA is transparent: 400 high-quality KonIQ-10k images are synthetically degraded (downsampling, brightness reduction, Gaussian noise) and then enhanced by nine AI models, followed by MOS annotation and evaluation of existing IQA models. There is no fitted-parameter derivation or 'prediction' that is statistically forced by the inputs; model evaluations use held-out human MOS and external models trained on other datasets. The central contribution, the dataset itself, is self-contained and not circular. However, the interpretative claim that AI-UGC perceptual quality is 'more similar to UGC' is partially endogenized by the dataset's construction, because AI-UGC is defined and built from UGC source images. This makes the conclusion somewhat self-definitional, though it is not a formal circular derivation. Self-citations appear in the references, but none are load-bearing for the main results—they are baselines or related work. Overall, the circularity is minor and limited to the framing of the conclusion, not the core benchmark construction.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

No fitted model parameters or invented physical entities. The dataset design does require hand-chosen degradation severities (unreported values), and the central interpretation rests on two domain assumptions: the synthetic-proxy mapping to real AI-UGC, and the comparability of the UGC versus AIGC fine-tuning regimes. The 'concept of AI-UGC' is a category definition, not a new entity.

free parameters (1)
  • degradation severity parameters = not reported (downsampling scale, brightness reduction factor, Gaussian noise sigma)
    Chosen by hand in Section 3.2 to create the low-resolution, low-light, and noisy inputs; they shape the MOS distributions and thus what 'AI-UGC' means in this benchmark, but their values are never stated.
assumptions (4)
  • domain assumption Synthetic degradation of clean, high-quality KonIQ-10k photos followed by AI enhancement reproduces the perceptual characteristics of real-world AI-enhanced UGC.
    Section 3.2 adopts this proxy because real UGC 'lack standardized degradation types'; the benchmark's external validity rests on it.
  • domain assumption Mean Opinion Score from five trained annotators, with an informal consistency cross-check, is a reliable quality label without reported inter-rater agreement statistics.
    Section 3.2 describes the annotation but gives no ICC, Krippendorff alpha, or score standard deviations.
  • domain assumption Fine-tuning an LMM on KADID-10K (about 10k images) yields a 'UGC quality assessment model' comparable in strength to one fine-tuned on AGIQA-3K (3k images) for 'AIGC quality assessment'.
    Section 4.1 defines the two model categories this way; the headline comparison depends on the two fine-tuning sets being comparably potent.
  • standard math PLCC and SRCC are the accepted evaluation metrics for IQA benchmarking.
    Used in Tables 1-2 to rank models; standard in the IQA literature.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AU-IQA: A Benchmark Dataset for Perceptual Quality Assessment of AI-Enhanced User-Generated Content." pith.science (2026). https://pith.science/paper/2FWJAKQE

@misc{pith2026250805016,
  author       = {Pith},
  title        = {Pith review of: AU-IQA: A Benchmark Dataset for Perceptual Quality Assessment of AI-Enhanced User-Generated Content},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2FWJAKQE}},
  note         = {Machine review of arXiv:2508.05016}
}
read the original abstract

AI-based image enhancement techniques have been widely adopted in various visual applications, significantly improving the perceptual quality of user-generated content (UGC). However, the lack of specialized quality assessment models has become a significant limiting factor in this field, limiting user experience and hindering the advancement of enhancement methods. While perceptual quality assessment methods have shown strong performance on UGC and AIGC individually, their effectiveness on AI-enhanced UGC (AI-UGC) which blends features from both, remains largely unexplored. To address this gap, we construct AU-IQA, a benchmark dataset comprising 4,800 AI-UGC images produced by three representative enhancement types which include super-resolution, low-light enhancement, and denoising. On this dataset, we further evaluate a range of existing quality assessment models, including traditional IQA methods and large multimodal models. Finally, we provide a comprehensive analysis of how well current approaches perform in assessing the perceptual quality of AI-UGC. The access link to the AU-IQA is https://github.com/WNNGGU/AU-IQA-Dataset.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

62 extracted references · 53 canonical work pages

  1. [1]

    Lorenzo Agnolucci, Leonardo Galteri, Marco Bertini, and Alberto Del Bimbo

  2. [2]

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng X...

    arXiv 2023

  3. [3]

    Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. 2024. TOPIQ: A Top-down Approach from Semantics to Distortions for Image Quality Assessment. In IEEE Transactions on Image Processing

    work page 2024

  4. [4]

    Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. 2023. Masked Image Training for Generalizable Deep Image Denoising. In IEEE/CVF Conference on Computer Vision and Pattern Recognition

    work page 2023

  5. [5]

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. 2024. Internvl: Scaling up Vision Foundation Models and Aligning for Generic Visual-linguistic Tasks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition

    work page 2024

  6. [6]

    Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang. 2016. Image Super-Resolution Using Deep Convolutional Networks. In IEEE Transactions on Pattern Analysis and Machine Intelligence

    work page 2016

  7. [7]

    Wei Dong, Han Zhou, Yulun Zhang, Xiaohong Liu, and Jun Chen. 2024. ECMamba: Consolidating Selective State Space Model with Retinex Guidance for Efficient Multiple Exposure Correction. In Advances in Neural Information Processing Systems

    work page 2024

  8. [8]

    Yunlong Dong, Xiaohong Liu, Yixuan Gao, Xunchu Zhou, Tao Tan, and Guangtao Zhai. 2023. Light-VQA: A Multi-Dimensional Quality Assessment Model for Low-Light Video Enhancement. In ACM International Conference on Multimedia

    work page 2023

Show all 62 references
  1. [9]

    Kang Fu, Yicong Peng, Zicheng Zhang, Qihang Xu, Xiaohong Liu, Jia Wang, and Guangtao Zhai. 2024. Attentionlut: Attention Fusion-based Canonical Polyadic Lut for Real-time Image Enhancement. In IEEE International Conference on Acous- tics, Speech and Signal Processing

    2024

  2. [10]

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative adversarial nets. In Advances in Neural Information Processing Systems

    2014

  3. [11]

    Jiang Hai, Binhao Guan, Zhen Liu, Xiaohong Liu, Jian Yu, Zheng Liu, Songchen Han, and Shuaicheng Liu. 2025. Learning to See in the Extremely Dark. In arXiv preprint arXiv:2506.21132

    2025 arXiv

  4. [12]

    Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. 2020. KonIQ-10k: An Ecologically Valid Database for Deep Learning of Blind Image Quality Assessment. In IEEE Transactions on Image Processing

    2020

  5. [13]

    Qiang Hu, Qihan He, Houqiang Zhong, Guo Lu, Xiaoyun Zhang, Guangtao Zhai, and Yanfeng Wang. 2025. VARFVV: View-Adaptive Real-Time Interactive Free- View Video Streaming with Edge Computing. In IEEE Journal on Selected Areas in Communications

    2025

  6. [14]

    Hai Jiang, Ao Luo, Xiaohong Liu, Songchen Han, and Shuaicheng Liu. 2024. LightenDiffusion: Unsupervised Low-Light Image Enhancement with Latent- Retinex Diffusion Models. In European Conference on Computer Vision

    2024

  7. [15]

    Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. 2021. MUSIQ: Multi-scale Image Quality Transformer. In IEEE/CVF international conference on computer vision

    2021

  8. [16]

    Diederik P Kingma and Max Welling. 2013. Auto-encoding Variational Bayes. In arXiv preprint arXiv:1312.6114

    2013 arXiv

  9. [17]

    Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. 2023. Pick-a-Pic: An Open Dataset of User Preferences for Text-to- Image Generation. In arXiv preprint arXiv:2305.01569

    2023 arXiv

  10. [18]

    Tengchuan Kou, Xiaohong Liu, Zicheng Zhang, Chunyi Li, Haoning Wu, Xiongkuo Min, Guangtao Zhai, and Ning Liu. 2024. Subjective-Aligned Dateset and Metric for Text-to-Video Quality Assessment. In ACM International Confer- ence on Multimedia

    2024

  11. [19]

    Chunyi Li, Tengchuan Kou, Yixuan Gao, Yuqin Cao, Wei Sun, Zicheng Zhang, Yingjie Zhou, Zhichao Zhang, Weixia Zhang, Haoning Wu, et al. 2024. AIGIQA- 20k: A Large Database for AI-Generated Image Quality Assessment. In arXiv preprint arXiv:2404.03407

    2024 arXiv

  12. [20]

    Chunyi Li, Zicheng Zhang, Haoning Wu, Wei Sun, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. 2023. AGIQA-3k: An Open Database For AI-Generated Image Quality Assessment. In IEEE Transactions on Circuits and Systems for Video Technology

    2023

  13. [21]

    Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. 2024. LLaVA-NeXT-Interleave: Tackling Multi-image, Video, and 3D in Large Multimodal Models. In arXiv preprint arXiv:2407.07895

    2024 arXiv

  14. [22]

    Wenhao Li, Guangyang Wu, Wenyi Wang, Peiran Ren, and Xiaohong Liu

  15. [23]

    Xinyue Li, Huiyu Duan, Jia Wang, Xiaohong Liu, Yitong Chen, and Guangtao Zhai. 2025. Situation-adaptive Neural Network for Fast Pre-computing Image Enhancement. In Science China Information Sciences

    2025

  16. [24]

    Hanhe Lin, Vlad Hosu, and Dietmar Saupe. 2019. KADID-10k: A Large-scale Artificially Distorted IQA Database. In International Conference on Quality of Multimedia Experience

    2019

  17. [25]

    Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Wanli Ouyang, Yu Qiao, and Chao Dong. 2024. DiffBIR: Towards Blind Image Restora- tion with Generative Diffusion Prior. InEuropean Conference on Computer Vision

    2024

  18. [26]

    Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjölund, and Thomas B Schön

  19. [27]

    Nikolay Ponomarenko, Lina Jin, Oleg Ieremeiev, Vladimir Lukin, Karen Egiazar- ian, Jaakko Astola, Benoit Vozel, Kacem Chehdi, Marco Carli, Federica Battisti, and C.-C. Jay Kuo. 2013. Image database TID2013: Peculiarities, results and perspectives. In Signal Processing: Image C...

    2013

  20. [28]

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen

  21. [29]

    In IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Refusion: Enabling Large-Size Realistic Image Restoration with Latent- Space Diffusion Models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition

  22. [30]

    Lee, Jonathan Ho, Tim Salimans, David J

    Chitwan Saharia, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. 2021. Palette: Image-to-Image Diffusion Models. In arXiv preprint arXiv:2111.05826

    2021 arXiv

  23. [31]

    Zhihao Shi, Xiaohong Liu, Kangdi Shi, Linhui Dai, and Jun Chen. 2024. Video frame interpolation via generalized deformable convolution. In IEEE transactions on multimedia

    2024

  24. [32]

    Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. 2020. Blindly Assess Image Quality in the Wild Guided by A Self-Adaptive Hyper Network. In IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion

    2020

  25. [33]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution Image Synthesis with Latent Diffusion Models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2022

  26. [34]

    Hossein Talebi and Peyman Milanfar. 2018. NIMA: Neural Image Assessment. In IEEE Transactions on Image Processing

    2018

  27. [35]

    Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. 2023. Exploring CLIP for Assessing the Look and Feel of Images. In AAAI Conference on Artificial Intelligence

    2023

  28. [36]

    Puyi Wang, Wei Sun, Zicheng Zhang, Jun Jia, Yanwei Jiang, Zhichao Zhang, Xiongkuo Min, and Guangtao Zhai. 2024. Large Multi-modality Model Assisted AI-Generated Image Quality Assessment. In ACM International Conference on Multimedia

    2024

  29. [37]

    Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. 2020. Blindly Assess Image Quality in the Wild Guided by a Self-Adaptive Hyper Network. In IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion

    2020

  30. [38]

    Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. 2018. ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks. In European Conference on Computer Vision Workshops

    2018

  31. [39]

    Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. 2022. Uformer: A General U-Shaped Transformer for Image Restoration. In IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2022

  32. [40]

    Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Chunyi Li, Liang Liao, Annan Wang, Erli Zhang, Wenxiu Sun, Qiong Yan, Xiongkuo Min, Guangtai Zhai, and Weisi Lin. 2024. Q-Align: Teaching LMMs for Visual Scoring via Discrete Text-Defined Levels. In International Conferen...

    2024

  33. [41]

    Wenjing Wang, Huan Yang, Jianlong Fu, and Jiaying Liu. 2024. Zero-Reference Low-Light Enhancement via Physical Quadruple Priors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2024

  34. [42]

    Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. 2024. One-Step Effective Diffusion Network for Real-World Image Super-Resolution. InAdvances in Neural Information Processing Systems

    2024

  35. [43]

    Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. 2023. Human Preference Score v2: A Solid Benchmark for Evaluating Human Preferences of Text-to-Image Synthesis. In arXiv preprint arXiv:2306.09341

    2023 arXiv

  36. [44]

    Shuzhou Yang, Moxuan Ding, Yanmin Wu, Zihan Li, and Jian Zhang. 2023. Im- plicit Neural Representation for Cooperative Low-light Image Enhancement. In IEEE/CVF International Conference on Computer Vision

    2023

  37. [45]

    Haoning Wu, Hanwei Zhu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, and Chunyi Li. 2024. Towards Open-ended Visual Quality Comparison. In European Conference on Computer Vision

    2024

  38. [46]

    Zhiyuan You, Xin Cai, Jinjin Gu, Tianfan Xue, and Chao Dong. 2025. Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution. In IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2025

  39. [47]

    Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. 2024. Scaling Up to Excellence: Practicing Model Scaling for Photo-Realistic Image Restoration In the Wild. In IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2024

  40. [48]

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. 2022. Restormer: Efficient Transformer for High-Resolution Image Restoration. In IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2022

  41. [49]

    Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. 2024. Pixel- Aware Stable Diffusion for Realistic Image Super-Resolution and Personalized �� ���� ������� ������ ����� ������� �������� ������ ���� �� ��� Stylization. In European Conference on Computer Vision

    2024

  42. [50]

    Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang. 2018. Blind Image Quality Assessment Using A Deep Bilinear Convolutional Neural Network. In IEEE Transactions on Circuits and Systems for Video Technology

    2018

  43. [51]

    Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. 2023. Blind Image Quality Assessment via Vision-Language Correspondence: A Mul- titask Learning Perspective. In IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2023

  44. [52]

    Zicheng Zhang, Tengchuan Kou, Shushi Wang, Chunyi Li, Wei Sun, Wei Wang, Xiaoyu Li, Zongyu Wang, Xuezhi Cao, Xiongkuo Min, Xiaohong Liu, and Guang- tao Zhai. 2025. Q-Eval-100K: Evaluating Visual Quality and Alignment Level for Text-to-Vision Content. In IEEE/CVF Conference on ...

    2025

  45. [53]

    Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. 2017. Be- yond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising. In IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2017

  46. [54]

    Zicheng Zhang, Yingjie Zhou, Chunyi Li, Kang Fu, Wei Sun, Xiaohong Liu, Xiongkuo Min, and Guangtao Zhai. 2024. A Reduced-Reference Quality Assess- ment Metric for Textured Mesh Digital Humans. In International Conference on Acoustics, Speech, and Signal Processing

    2024

  47. [55]

    Zicheng Zhang, Yingjie Zhou, Chunyi Li, Baixuan Zhao, Xiaohong Liu, and Guangtao Zhai. 2024. Quality Assessment in the Era of Large Models: A Survey. In ACM Transactions on Multimedia Computing, Communications and Applications

    2024

  48. [56]

    Han Zhou, Wei Dong, Xiaohong Liu, Shuaicheng Liu, Xiongkuo Min, Guangtao Zhai, and Jun Chen. 2024. GLARE: Low Light Image Enhancement via Generative Latent Feature based Codebook Retrieval. In European Conference on Computer Vision

    2024

  49. [57]

    Zicheng Zhang, Chunyi Li, Wei Sun, Xiaohong Liu, Xiongkuo Min, and Guangtao Zhai. 2023. A Perceptual Quality Assessment Exploration for AIGC Images. In IEEE International Conference on Multimedia and Expo Workshops

    2023

  50. [58]

    Yingjie Zhou, Zicheng Zhang, Wei Sun, Xiaohong Liu, Xiongkuo Min, Zhihua Wang, Xiaoping Zhang, and Guangtao Zhai. 2024. THQA: A Perceptual Quality Assessment Database for Talking Heads. In IEEE International Conference on Image Processing

    2024

  51. [61]

    Xunchu Zhou, Xiaohong Liu, Yunlong Dong, Tengchuan Kou, Yixuan Gao, Zicheng Zhang, Chunyi Li, Haoning Wu, and Guangtao Zhai. 2024. Light-VQA+: A Video Quality Assessment Model for Exposure Correction with Vision-Language Guidance. In arXiv preprint arXiv:2405.03333

    2024 arXiv

  52. [2022]

    In arXiv preprint arXiv:2204.06125

    Hierarchical Text-Conditional Image Generation with Clip Latents. In arXiv preprint arXiv:2204.06125

    arXiv

  53. [2023]

    In ACM International Conference on Multimedia

    FastLLVE: Real-Time Low-Light Video Enhancement with Intensity-Aware Lookup Table. In ACM International Conference on Multimedia

  54. [2024]

    In IEEE/CVF Winter Conference on Applications of Computer Vision

    ARNIQA: Learning Distortion Manifold for Image Quality Assessment. In IEEE/CVF Winter Conference on Applications of Computer Vision

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.