Pith. sign in

REVIEW 3 major objections 7 minor 3 cited by

IE-Bench: Advancing the Measurement of Text-Driven Image Editing for Human Perception Alignment

T0 review · 3 major / 7 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A new benchmark and metric for text-driven image editing aim to score edited images the way human raters do.

desk verdict Useful dataset idea and a plausible metric design, but the sample-level 10-fold CV lets the source branch memorize source identity, so the headline gain over IP-IQA is likely inflated. read the letter →

arxiv 2501.09927 v1 pith:ZAW3TLKQ submitted 2025-01-17 cs.CV cs.AI

Shangkun Sun , Bowen Qu , Xiaoyu Liang , Songlin Fan , Wei Gao This is my paper · ORCID
classification cs.CVcs.AI
keywords text-drivenimageeditingbenchmarkqualityassessmentmeanopinionscorehumanperceptionalignmentsource-awaremetricmulti-modalevaluationCLIP-basedscoring
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 sets out to close a gap in evaluating text-driven image editing: existing metrics look at text-image alignment or low-level similarity, but none model how the edited image should relate to the source image while also matching what humans perceive. The authors build IE-Bench, a benchmark suite whose database IE-DB contains 301 source images, edited results from five editing methods, and 3,010 mean opinion scores collected from 25 human raters. On top of it they train IE-QA, a source-aware multi-modal quality assessment network that scores an edit from the source image, the edited image, and the prompt together. Their central claim is that IE-QA aligns with human subjective ratings far better than previous objective and learned metrics, with gains of about 10 percentage points in correlation on their benchmark. If correct, the field gains a reproducible, human-aligned yardstick for comparing editing models and for training better ones.

What carries the argument

IE-QA's architecture is the mechanism that carries the argument. It uses a frozen CLIP visual encoder for the edited image and a CLIP text encoder for the prompt, feeding their interaction through cross-modality attention pooling to judge text-image alignment; a second branch encodes the source and edited images and concatenates their features through a lightweight feed-forward network to score the source-target relationship; and a third branch inherits the visual quality evaluation from IP-IQA, assessing aesthetics and technical distortion. The three scores are fused in a head trained with a Pearson correlation loss plus a rank loss. The ablation study shows that removing either the text branch or the source branch degrades correlation, and that simple concatenation beats attention-based fusion, supporting the claim that the improvement is not just added parameters.

What would settle it

Re-run the subjective experiment with an independent panel on the same 301 source images and prompts, then recompute IE-QA's SROCC and PLCC against the new MOS. If the gap over IP-IQA shrinks below chance or fails to reproduce, the human-alignment claim collapses. Alternatively, compute inter-rater agreement on IE-DB directly: low agreement would mean the MOS itself is not a reliable target.

Watch

Extended reading notes

Core claim

The paper's central claim is that IE-Bench is the first image quality assessment dataset and model built specifically for text-driven image editing, and that IE-QA is the first method to align with human perception on this task by explicitly modeling the source-target relationship. On the IE-DB benchmark, IE-QA reports SROCC 0.7520, PLCC 0.7498, KRCC 0.5541, and RMSE 1.045, outperforming prior VQA methods and objective metrics by 10.46%, 9.02%, 10.19%, and 11.08% in those four metrics. Compared with its strongest baseline IP-IQA, the gains are 0.1396, 0.1435, 0.096, and 0.243 respectively. The authors argue this demonstrates that modeling the dynamic connection between source and edited image, conditioned on the prompt, is the missing ingredient in current editing metrics.

Load-bearing premise

The load-bearing premise is that the 3,010 mean opinion scores from 25 raters are a stable, unbiased ground truth for edit quality; the paper reports no inter-rater agreement, per-item variance, or power analysis, so if those ratings are noisy the measured correlations, including IE-QA's advantage, are not a stable yardstick.

Editorial extensions

If this is right

  • Editing models can be compared on a shared MOS-labeled benchmark instead of one-off subjective experiments, making results reproducible across labs.
  • IE-QA can serve as a drop-in evaluation metric for future text-driven editing papers, complementing or replacing CLIP-score-based reporting.
  • Metrics that ignore the source image will systematically misrank edits that should preserve identity, such as removing an accessory or changing style.
  • The explicit source-target branch could be adapted to assess instruction-following in other conditional generation tasks where input and output share structure.

Reading between the lines

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

  • The reported correlation gain is measured on a 301-image benchmark with 25 raters; whether the advantage holds on larger, more heterogeneous editing corpora is untested, so the metric's rank among editing evaluators should be read as provisional.
  • If IE-QA's design transfers, the same source-aware text-conditioned scoring could be used as a reward signal for fine-tuning editing models, turning an evaluator into a training objective.
  • The paper does not report inter-rater agreement; a replication study that computes per-item variance or Krippendorff's alpha would test the stability of the MOS ground truth that all correlations rely on.
  • Because prompts were manually designed per image, the benchmark measures performance on human-curated instructions; automatically generated prompt distributions might rank the same editing models differently.
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 / 7 minor

Summary. The paper introduces IE-Bench, a benchmark suite for text-driven image editing quality assessment, comprising IE-DB (a dataset of 301 source images spanning real-world, CG, AIGC, and art; edited by five methods under manually designed prompts; rated by 25 subjects using Z-score normalization and BT.500 outlier filtering) and IE-QA (a multi-modal, source-aware quality assessment network that combines a CLIP-based text-image alignment branch, a source-target relationship branch via feature concatenation, and an IP-IQA-style visual quality branch, trained with a PLCC and rank loss). On 10-fold cross-validation over IE-DB, IE-QA reports SROCC 0.7520, PLCC 0.7498, KRCC 0.5541, and RMSE 1.045, outperforming the strongest baseline IP-IQA (SROCC 0.6124) and other objective and human-aligned metrics. The authors claim this is the first IQA dataset and model tailored to text-driven image editing.

Significance. If the reported results are robust, IE-Bench would be a valuable first step: a MOS-labeled dataset for text-driven image editing, and IE-QA would provide a reference method that explicitly models the source-target relationship. The authors are to be credited for the manual curation of diverse source images, recruitment of 25 participants with varied backgrounds following ITU guidelines, application of Z-score normalization and BT.500 outlier screening, and for an 'additional parameters' ablation that attempts to separate design choices from model capacity. However, the significance is currently limited by three factors: the evaluation is an in-distribution 10-fold CV on a single manually curated dataset with no external validation; the dataset and code are promised but not released; and no reliability statistics or significance tests are reported for the subjective scores or the metric comparisons. The central quantitative claim is therefore plausible but not yet established.

major comments (3)
  1. [Section 5.1, Eqs. (4)–(6), Table 3] The 10-fold cross-validation in Section 5.1 is not described as grouped by source image. IE-DB contains multiple edited outputs per source image (five editing models, and prompt statistics in Section 3.2 indicate multiple prompts per image). A random sample-level split places different edits of the same source image in both training and validation folds. Because IE-QA concatenates source-image features with edited-image features (Eqs. (4)–(6)), the network can identify a source image seen during training and predict its average MOS rather than evaluating the specific edit. The 'Additional Parameters' ablation in Table 3 does not control for this: it matches parameter count but does not remove source identity from the input. To support the central claim that IE-QA achieves SROCC 0.7520 versus 0.6124 for IP-IQA, the authors must report results with folds grouped by source image (and ideally by source–prompt pair), including fold-wise standard deviations and a paired significance test.
  2. [Abstract and Section 3.4] The dataset description is internally inconsistent with respect to the number of opinion scores. The Abstract states 'total 3,010 Mean Opinion Scores (MOS) provided by 25 human subjects,' while Section 3.4 says each participant evaluated all image samples. With 301 source images and five editing methods (Section 3.3), there are 1,505 edited images; 3,010 is exactly twice that number but not 25 times it. The paper does not clarify what 3,010 counts (aggregated MOS per image after outlier filtering, ratings per participant, or a subset of the full rating matrix). Nor does it report inter-rater agreement, per-item variance, or the fraction of outliers removed by the BT.500 procedure. Because every correlation coefficient in Table 2 is computed against this ground truth, the unresolved bookkeeping and missing reliability statistics make the reported numerical results unverifiable.
  3. [Section 5.3, Table 2] The superiority claim is not quantitatively supported as reported. The paper states that IE-QA achieves improvements of '10.46%, 9.02%, 10.19%, 11.08%' in SROCC, PLCC, KLCC, and RMSE, but these numbers do not match the values in Table 2: for SROCC, 0.7520 versus 0.6124 is a relative improvement of about 22.8%. If the percentages are averages over a different comparison set, the definition should be given. More importantly, no confidence intervals, per-fold results, or significance tests are provided for any of the comparisons; the observed advantage over IP-IQA could be within sampling noise for a dataset of this size. Please report fold-wise statistics and a paired significance test (e.g., Wilcoxon signed-rank over folds).
minor comments (7)
  1. [Section 5.2] The text states that four evaluation metrics are used but lists only SROCC, PLCC, and KRCC; RMSE appears in Table 2 and should be added to the list.
  2. [Section 5.3] The abbreviation 'KLCC' should be 'KRCC' (Table 2 header correctly uses KRCC).
  3. [Table 2] The category 'Subjective Metrics' contains objective zero-shot metrics (CLIP-T, CLIP-V, DINO, LPIPS, SSIM, PSNR); rename the category to 'Zero-shot Objective Metrics' or similar.
  4. [Section 6] The conclusion refers to 'an extensive array of video content'; since IE-DB contains images, this should read 'image content'.
  5. [References] References [11] and [12] are the same work; the duplicate entry should be removed or merged.
  6. [Page 1 footnote] The footnote contains a typo: 'Corrsponding' should be 'Corresponding'.
  7. [Abstract] The authors state that all data and code will be made public, but no release URL or timetable is given; for a benchmark paper, providing a release link in the abstract would aid reproducibility.

Circularity Check

1 steps flagged · score 4.0 of 10

IE-QA's 10-fold evaluation on IE-DB is vulnerable to source-image leakage because folds are not stated to be grouped by source; the IP-IQA baseline is same-group work, though the architecture has independent content.

  1. fitted input called prediction [Section 5.1 Implementation Details; Section 4.3 Eq. (6); Table 2]
    "Following the 10-fold cross-validation approach used in previous studies [24,29,43], all models were trained on the IE DB dataset with an initial learning rate of 1e-3, a batch size of 8, and a total of 60 epochs. ... os = Hs(Concat(f, f⋆)) (6) ... IE-QA outperforms previous traditional VQA methods and commonly used objective metrics in aligning with human subjective perception, achieving improvements of 10.46%, 9.02%, 10.19%, 11.08% in SROCC, PLCC, KLCC, and RMSE, respectively."

    IE-DB contains 301 source images, each edited by five models (Sec. 3.1, Table 1), so many edited images share the exact same source tensor. IE-QA explicitly concatenates source-image features with edited-image features (Eq. 6). The 10-fold protocol is described only at the edited-sample level and never states that folds are grouped by source image. Under a sample-level split, every validation fold contains edits whose source image appeared in the training folds; the network can identify that source and regress toward a memorized source-average MOS rather than assessing the editing operation.

full rationale

IE-QA is a supervised regressor trained on IE-DB MOS and evaluated by 10-fold cross-validation on the same database; this is a standard, non-circular way to demonstrate within-distribution generalization, and no equation defines the predicted score as the MOS. The text-image branch and source-target branch are motivated by the editing task itself, not by fitting the evaluation scores. However, the evaluation protocol has a genuine leakage risk: because IE-DB has only 301 source images with multiple edited outputs per source, and IE-QA's source branch consumes the source image, a sample-level fold split lets the model memorize source identity, making some 'predictions' a lookup of training labels rather than an edit-quality judgment. The paper does not state that folds are grouped by source. In addition, the main baseline IP-IQA (Ref. [24]) is prior work by two of the present authors (Bowen Qu, Wei Gao) and is used as the visual-quality backbone of IE-QA, so the headline comparison is partly same-group; this self-citation is not load-bearing for the architectural derivation but reduces the independence of the empirical anchor. These issues justify a moderate circularity score, while the central claim retains independent content.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The benchmark claim rests on subject-rated ground truth and on CLIP-based features; these are domain assumptions rather than derived results. The hyperparameters listed are conventional choices, but alpha and the training schedule are hand-set.

free parameters (3)
  • Loss weight alpha = 0.3
    Eq. (7) sets L = L_plcc + alpha * L_rank with alpha = 0.3 'in practice'; the value is hand-chosen, not derived.
  • Input resolution = shorter side 512 px
    Section 3.1 resizes all source images to shorter side 512; this preprocessing choice affects all features.
  • Training schedule = 40 epochs linear probe + 20 epochs full fine-tune
    Section 5.1; chosen following IP-IQA, not derived from data.
assumptions (4)
  • domain assumption MOS from 25 subjects, after Z-score normalization and BT.500 outlier filtering, is a reliable ground truth for editing quality.
    Section 3.4; all correlations are computed against this ground truth, but no inter-rater agreement or variance is reported.
  • domain assumption The three rated dimensions (text-image consistency, source-target fidelity, overall quality) capture what humans mean by editing quality.
    Section 3.4 defines the rating criteria; the validity of the benchmark depends on this decomposition.
  • domain assumption CLIP visual and text features contain sufficient signal for text-image alignment and source-target relevance.
    Sections 4.2 and 4.3 build the entire model on CLIP encoders; no independent feature source is tested.
  • domain assumption The five selected editing models produce a quality distribution representative of text-driven image editing outputs.
    Section 3.3 selects five models across SD versions and paradigms; the benchmark's coverage depends on this choice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of IE-Bench: Advancing the Measurement of Text-Driven Image Editing for Human Perception Alignment." pith.science (2026). https://pith.science/paper/ZAW3TLKQ

@misc{pith2026250109927,
  author       = {Pith},
  title        = {Pith review of: IE-Bench: Advancing the Measurement of Text-Driven Image Editing for Human Perception Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZAW3TLKQ}},
  note         = {Machine review of arXiv:2501.09927}
}
read the original abstract

Recent advances in text-driven image editing have been significant, yet the task of accurately evaluating these edited images continues to pose a considerable challenge. Different from the assessment of text-driven image generation, text-driven image editing is characterized by simultaneously conditioning on both text and a source image. The edited images often retain an intrinsic connection to the original image, which dynamically change with the semantics of the text. However, previous methods tend to solely focus on text-image alignment or have not aligned with human perception. In this work, we introduce the Text-driven Image Editing Benchmark suite (IE-Bench) to enhance the assessment of text-driven edited images. IE-Bench includes a database contains diverse source images, various editing prompts and the corresponding results different editing methods, and total 3,010 Mean Opinion Scores (MOS) provided by 25 human subjects. Furthermore, we introduce IE-QA, a multi-modality source-aware quality assessment method for text-driven image editing. To the best of our knowledge, IE-Bench offers the first IQA dataset and model tailored for text-driven image editing. Extensive experiments demonstrate IE-QA's superior subjective-alignments on the text-driven image editing task compared with previous metrics. We will make all related data and code available to the public.

Figures

Figures reproduced from arXiv: 2501.09927 by the authors.

Figure 1
Figure 1. Overview of the proposed IE-Bench. works [7], we manually design diverse editing instructions for each image, covering aspects such as structural changes (e.g., shape, size), style changes (e.g., texture, color), and semantic changes (e.g., pose, action, addition, replacement, deletion). We then apply multiple methods to generate di￾verse edited results. Finally, we assemble 25 human partic￾ipants from various backg… view at source ↗
Figure 2
Figure 2. Collection of source images. (a) Sources of images. (b) Categories of images. (c) Content of images. (d) (f) denotes fine-grained [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Statistics of IE-DB prompts. (a) Word cloud of IE-Bench DB prompts. (b) Proportion of different types [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Z-score MOS distributions of different editing methods. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Network architecture of IE-QA. feature ebv is interacted with the text encoder through cross￾attention, denoted as Fca. 4.3. Source-Target Relationship Assessing the consistency between original and edited im￾ages is inherently complex due to their latent connections a…
Figure 6
Figure 6. Figure 6: Demo of scores in IE-QA. 5.4. Qualitative Results We further conducted a qualitative comparison for differ￾ent score levels in IE, as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

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

  1. DSH-Bench: A Difficulty- and Scenario-Aware Benchmark with Hierarchical Subject Taxonomy for Subject-Driven Text-to-Image Generation

    cs.CV 2026-03 unverdicted novelty 7.0 of 10

    DSH-Bench supplies a hierarchical 58-category subject set, difficulty/scenario labels, and a human-aligned SICS metric that exposes systematic failures of 19 subject-driven T2I models.

  2. LMM4Edit: Benchmarking and Evaluating Multimodal Image Editing with LMMs

    cs.CV 2025-07 conditional novelty 7.0 of 10

    A large human-annotated benchmark of AI-edited images (EBench-18K) plus a fine-tuned LMM metric (LMM4Edit) that predicts human preference scores across three dimensions and answers editing-specific questions.

  3. ADIEE: Automatic Dataset Creation and Scorer for Instruction-Guided Image Editing Evaluation

    cs.CV 2025-07 conditional novelty 6.0 of 10

    An automatically generated training dataset and a fine-tuned LLaVA-NeXT model produce an image editing evaluation scorer that aligns with human preference and serves as a reward model for improving editing models.

Reference graph

Works this paper leans on

44 extracted references · 31 canonical work pages · cited by 3 Pith papers

  1. [1]

    In- structpix2pix: Learning to follow image editing instructions

    Tim Brooks, Aleksander Holynski, and Alexei A Efros. In- structpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392–18402, 2023. 1, 3, 4, 5

    work page 2023

  2. [2]

    Masactrl: Tuning-free mu- tual self-attention control for consistent image synthesis and editing

    Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xi- aohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mu- tual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 22560–22570, 2023. 1, 3, 4, 5

    work page 2023

  3. [3]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural informa- tion processing systems, 34:8780–8794, 2021. 3

    2021

  4. [4]

    Gpt-3: Its nature, scope, limits, and consequences

    Luciano Floridi and Massimo Chiriatti. Gpt-3: Its nature, scope, limits, and consequences. Minds and Machines, 30: 681–694, 2020. 3

    2020

  5. [5]

    Gans trained by a two time-scale update rule converge to a local nash equilib- rium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- rium. Advances in neural information processing systems , 30, 2017. 3, 7

    work page 2017

  6. [6]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020. 3

    2020

  7. [8]

    Diffusion model-based image editing: A survey

    Yi Huang, Jiancheng Huang, Yifan Liu, Mingfu Yan, Jiaxi Lv, Jianzhuang Liu, Wei Xiong, He Zhang, Shifeng Chen, and Liangliang Cao. Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525, 2024. 4

    arXiv 2024

  8. [9]

    Smartedit: Exploring complex instruction-based image editing with multimodal large lan- guage models

    Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, et al. Smartedit: Exploring complex instruction-based image editing with multimodal large lan- guage models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 8362– 8371, 2024. 3, 4

    work page 2024

Show all 44 references
  1. [10]

    Methodology for the subjective as- sessment of the quality of television pictures itu-r recommen- dation

    Int.Telecommun.Union. Methodology for the subjective as- sessment of the quality of television pictures itu-r recommen- dation. Tech. Rep., 2000. 5

    2000

  2. [12]

    Convolu- tional neural networks for no-reference image quality assess- ment

    Le Kang, Peng Ye, Yi Li, and David Doermann. Convolu- tional neural networks for no-reference image quality assess- ment. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1733–1740, 2014. 2

    2014

  3. [13]

    The tum high definition video datasets

    Christian Keimel, Arne Redl, and Klaus Diepold. The tum high definition video datasets. In 2012 Fourth international workshop on quality of multimedia experience , pages 97–

    2012

  4. [14]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 ,

    arXiv

  5. [15]

    Pick-a-pic: An open dataset of user preferences for text-to-image generation

    Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Ma- tiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems , 36: 36652–36663, 2023. 1, 3, 7, 8

    2023

  6. [16]

    Subjective-aligned dateset and metric for text-to-video qual- ity assessment

    Tengchuan Kou, Xiaohong Liu, Zicheng Zhang, Chunyi Li, Haoning Wu, Xiongkuo Min, Guangtao Zhai, and Ning Liu. Subjective-aligned dateset and metric for text-to-video qual- ity assessment. arXiv preprint arXiv:2403.11956, 2024. 5, 7

    2024 arXiv

  7. [17]

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

    Eric C Larson and Damon M Chandler. Most apparent dis- tortion: full-reference image quality assessment and the role of strategy. Journal of electronic imaging , 19(1):011006– 011006, 2010. 2

    2010

  8. [18]

    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 Sys- tems for Video Technology, 2023. 2, 5

    2023

  9. [19]

    Aigiqa-20k: A large database for ai-generated image quality assessment

    Chunyi Li, Tengchuan Kou, Yixuan Gao, Yuqin Cao, Wei Sun, Zicheng Zhang, Yingjie Zhou, Zhichao Zhang, Weixia Zhang, Haoning Wu, et al. Aigiqa-20k: A large database for ai-generated image quality assessment. arXiv preprint arXiv:2404.03407, 2(3):5, 2024. 1

    2024 arXiv

  10. [20]

    Norm-in-norm loss with faster convergence and better performance for im- age quality assessment

    Dingquan Li, Tingting Jiang, and Ming Jiang. Norm-in-norm loss with faster convergence and better performance for im- age quality assessment. In Proceedings of the 28th ACM In- ternational conference on multimedia, pages 789–797, 2020. 2

    2020

  11. [21]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceeding...

    2014

  12. [22]

    Referring image editing: Object-level image editing via referring expressions

    Chang Liu, Xiangtai Li, and Henghui Ding. Referring image editing: Object-level image editing via referring expressions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13128–13138, 2024. 3

    2024

  13. [23]

    Dinov2: Learning robust visual features without super- vision

    Maxime Oquab, Timoth ´ee Darcet, Th´eo Moutakanni, Huy V V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without super- vision. Transactions on Machine Learning Research,...

    2024

  14. [24]

    Bringing textual prompt to ai-generated image quality assessment

    Bowen Qu, Haohui Li, and Wei Gao. Bringing textual prompt to ai-generated image quality assessment. arXiv preprint arXiv:2403.18714, 2024. 2, 6, 7, 8

    2024 arXiv

  15. [25]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

    2021

  16. [26]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 2250...

    2023

  17. [27]

    Methodology for the subjective assessment of the quality of television pictures

    B Series. Methodology for the subjective assessment of the quality of television pictures. Recommendation ITU-R BT , 500(13), 2002. 2, 5

    2002

  18. [28]

    Doubly abductive coun- terfactual inference for text-based image editing

    Xue Song, Jiequan Cui, Hanwang Zhang, Jingjing Chen, Richang Hong, and Yu-Gang Jiang. Doubly abductive coun- terfactual inference for text-based image editing. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9162–9171, 2024. 1, 3

    2024

  19. [29]

    Blindly assess image qual- ity 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 qual- ity in the wild guided by a self-adaptive hyper network. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 7, 8

    2020

  20. [30]

    E-bench: Subjective-aligned benchmark suite for text-driven video editing quality assessment

    Shangkun Sun, Xiaoyu Liang, Songlin Fan, Wenxu Gao, and Wei Gao. E-bench: Subjective-aligned benchmark suite for text-driven video editing quality assessment. arXiv preprint arXiv:2408.11481, 2024. 5

    2024 arXiv

  21. [31]

    Improved artgan for conditional synthesis of natural image and artwork

    Wei Ren Tan, Chee Seng Chan, Hernan Aguirre, and Kiyoshi Tanaka. Improved artgan for conditional synthesis of natural image and artwork. IEEE Transactions on Image Processing, 28(1):394–409, 2019. 4

    2019

  22. [32]

    Full-reference image quality as- sessment by combining features in spatial and frequency do- mains

    Zhisen Tang, Yuanlin Zheng, Ke Gu, Kaiyang Liao, Wei Wang, and Miaomiao Yu. Full-reference image quality as- sessment by combining features in spatial and frequency do- mains. IEEE Transactions on Broadcasting, 65(1):138–151,

  23. [33]

    Plug-and-play diffusion features for text-driven image-to-image translation

    Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 1921–1930, 2023. 3, 4, 5

    1921

  24. [34]

    Ex- ploring clip for assessing the look and feel of images

    Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Ex- ploring clip for assessing the look and feel of images. InPro- ceedings of the AAAI Conference on Artificial Intelligence , pages 2555–2563, 2023. 2

    2023

  25. [35]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600–612, 2004. 1, 3, 8

    2004

  26. [36]

    Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis

    Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341 ,

    arXiv

  27. [37]

    Human preference score: Better aligning text- to-image models with human preference

    Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hong- sheng Li. Human preference score: Better aligning text- to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 2096–2105, 2023. 1, 3, 7

    2023

  28. [38]

    Imagere- ward: Learning and evaluating human preferences for text- to-image generation, 2023

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagere- ward: Learning and evaluating human preferences for text- to-image generation, 2023. 1, 3, 8

    2023

  29. [39]

    Inversion-free image editing with natural language

    Sihan Xu, Yidong Huang, Jiayi Pan, Ziqiao Ma, and Joyce Chai. Inversion-free image editing with natural language. arXiv preprint arXiv:2312.04965, 2023. 5

    2023 arXiv

  30. [40]

    Magicbrush: A manually annotated dataset for instruction- guided image editing

    Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction- guided image editing. Advances in Neural Information Pro- cessing Systems, 36, 2024. 1, 3, 5

    2024

  31. [41]

    A comprehensive evaluation of full reference image quality as- sessment algorithms

    Lin Zhang, Lei Zhang, Xuanqin Mou, and David Zhang. A comprehensive evaluation of full reference image quality as- sessment algorithms. In 2012 19th IEEE International Con- ference on Image Processing, pages 1477–1480. IEEE, 2012. 2

    2012

  32. [42]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 586–595, 2018. 1, 3, 8

    2018

  33. [43]

    Blind image quality assessment using a deep bilinear convolutional neural network

    Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang. Blind image quality assessment using a deep bilinear convolutional neural network. IEEE Transactions on Cir- cuits and Systems for Video Technology, 30(1):36–47, 2018. 2, 3, 7, 8

    2018

  34. [44]

    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 IEEE Conference on Computer Vision and Pattern Recog- nition, pages 14071–14081, 2023. 2, 5

    2023

  35. [45]

    Sine: Single image editing with text- to-image diffusion models

    Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text- to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 6027–6037, 2023. 1, 3

    2023

  36. [46]

    Scene parsing through ade20k dataset

    Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017. 4

    2017

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.