Pith. sign in

REVIEW 4 major objections 4 minor 148 references

A new benchmark ranks 10 open-source image editors across 21 tasks and finds a clear gap between atomic and compositional editing.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 01:48 UTC pith:OA6KW5LF

load-bearing objection A useful survey with a new benchmark that is not yet verifiable: no data/code release, no inter-rater stats, and no leakage analysis between the curated test pairs and training sources of evaluated models. the 4 major comments →

arxiv 2607.25642 v1 pith:OA6KW5LF submitted 2026-07-28 cs.CV cs.CL

Instruction-based Image Editing: A Survey on Data, Models, Evaluation, and Applications

Xianghao Zang , Zijian Jiang , Jiarong Cheng , Qianrui Teng , Ying He , Yuxuan Mu , Chao Ban , Huayu Zhang
show 3 more authors
Lanxiang Zhou Zerun Feng Chi Zhang
This is my paper
classification cs.CV cs.CL
keywords instruction-based image editingCDD-IIE Benchbenchmarkdiffusion modelsautoregressive modelscompositional editingspatial reasoningevaluation metrics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper is a survey of instruction-based image editing (IIE), organized around task taxonomy, data construction, model architectures, and evaluation. Its central contribution is CDD-IIE Bench, a curated benchmark of 1,353 image-instruction pairs spanning 21 editing tasks grouped into two suites: basic atomic editing and advanced compositional editing. Twelve expert annotators scored each model output on 3 dimensions per task, and the averages place Qwen-Image-Edit-2509 and Uniworld-V2 at the top, followed by Step1X-Edit-v1.2 and Flux.2-dev. The authors draw three diagnostic conclusions: current models are strong at atomic edits but weaker at compositional reasoning, spatial understanding is the biggest bottleneck, and no model is uniformly good across all task types. The paper argues that such a structured, dual-suite evaluation can serve as a standardized measure of IIE capability.

Core claim

The paper proposes CDD-IIE Bench, a benchmark for Comprehensive, in-Depth, and Diagnostic evaluation of instruction-based image editing. It consists of 1,353 image-instruction pairs covering 21 editing tasks organized into five dimensions (object-level operations, image-level operations, utility tasks, complex instruction and reasoning, spatial understanding and reasoning) and two suites (atomic and compositional). Each task is scored by 12 experts on three tailored dimensions, with four experts per dimension. The empirical results show a performance hierarchy: Qwen-Image-Edit-2509 (3.72) and Uniworld-V2 (3.70) lead, followed by Step1X-Edit-v1.2 (3.61) and Flux.2-dev (3.53), with older model

What carries the argument

The central object is the CDD-IIE Bench dataset, built from curated pairs drawn from existing benchmarks (AnyEdit, GEdit-Bench, ImgEdit-Bench, CompBench, and others) plus self-collected internet pairs. Its evaluative power comes from a dual-suite structure: a basic atomic suite covering 13 tasks and an advanced compositional suite covering 8 tasks. Each task is assessed on three task-specific dimensions with an explicit 1–5 rubric, and the final score for each model is the average over dimensions and models. This design is meant to provide both breadth (comprehensiveness) and depth (diagnosability) in a compact, efficient evaluation set.

Load-bearing premise

The benchmark's 1,353 test pairs are assumed to be independent of the training data of the evaluated models; if any overlap exists, the rankings may reflect memorization rather than editing skill.

What would settle it

Running CDD-IIE Bench on the same 10 models after explicitly removing any source-benchmark pairs that appear in their training corpora would reveal whether the top performers' scores drop substantially when leakage is eliminated. A second check: recompute scores using a different panel of experts (or a larger panel) and compare the resulting ranking; if the top-two ordering changes, the current hierarchy is not stable under rater variation.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If CDD-IIE Bench is adopted as a standard, model comparisons in the field will be reported on a common 21-task, 5-dimension scale, making cross-paper ranking easier.
  • The observed atomic-vs-compositional gap implies that next-generation editors should be evaluated on spatial reasoning and multi-step instructions, not just single-object edits.
  • The low scores on utility tasks (e.g., image repair, visual effect removal) suggest that current training paradigms over-emphasize semantic editing at the expense of low-level fidelity, pointing toward multi-task training regimes.
  • The benchmark's diagnostic design could be reused to probe specific failure modes, such as viewpoint transformation and count change, which the results identify as persistent weak points.
  • Releasing the benchmark openly would allow third-party models to be tested on the same 1,353 pairs, enabling fairer longitudinal tracking of progress in IIE.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because CDD-IIE Bench draws its pairs from the very datasets that some evaluated models were trained on, the reported rankings may partly reflect train-test overlap rather than pure editing ability; the paper does not report leakage analysis.
  • The lack of reported inter-rater agreement means that the averaged expert scores, while detailed, may not be stable; a future version could publish variance or agreement statistics to strengthen the reliability of the hierarchy.
  • The benchmark's compact size (1,353 pairs) makes it feasible to run repeatedly, which could enable fast iterative development of models aimed at closing the spatial-reasoning gap.
  • The finding that spatial understanding is the universal bottleneck suggests that a targeted probe—e.g., a dedicated viewpoint- and count-change suite with synthetic ground truth—might yield more actionable diagnostics than the current broad rubric.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper is a survey of instruction-based image editing (IIE), organized around a taxonomy of 21 editing tasks in two suites (atomic and compositional), a review of data construction methods, model architectures (GAN-CLIP, diffusion, autoregressive, unified multimodal), evaluation metrics, and commercial systems. As its main empirical contribution, it introduces CDD-IIE Bench, a 1,353-pair benchmark curated from existing benchmarks plus internet-collected pairs, and reports human expert ratings of 10 open-source models on 21 tasks across 5 dimensions. The reported hierarchy places Qwen-Image-Edit-2509 and Uniworld-V2 at the top, followed by Step1X-Edit-v1.2 and Flux.2-dev, with older models lower. The paper claims the benchmark is comprehensive, in-depth, and diagnostic.

Significance. If the benchmark is valid, it would be a useful standardized evaluation resource for a fast-moving field: the 21-task, 5-dimension structure is more granular than most existing benchmarks, the dual atomic/compositional design supports diagnostic analysis, and the human-evaluation effort (1,353 pairs × 3 dimensions × 4 experts × 10 models = 162,360 ratings, which is arithmetically consistent) is substantial. The survey portion is broad and reasonably current, covering data pipelines, architectural evolution, and evaluation practice. However, the benchmark's central claim of being a rigorous, standardized measure rests on two premises that the manuscript does not establish: independence from the evaluated models' training data, and reliability of the human ratings. These are load-bearing for the headline ranking and diagnostic conclusions.

major comments (4)
  1. [§5.5.2, §5.7] The benchmark is not shown to be independent of the evaluated models' training data. The 1,353 pairs are curated from GEdit-Bench, AnyEdit, I2EBench, ImgEdit-Bench, CompBench, KRIS-Bench, ByteMorph, Reason-Edit, RefEdit-Bench, and ComplexBench-Edit, and all 1,353 pairs are then used to evaluate all models (§5.7). At least Step1X-Edit-v1.2 is from the same line of work as GEdit-Bench [22], and several other evaluated models may have been trained on subsets of these exact datasets (e.g., AnyEdit-derived data). The paper performs no overlap or leakage analysis and does not report per-source sample counts. Without this, the reported hierarchy may reflect train-test contamination rather than editing ability. The authors should report exact source splits, perform near-duplicate image/instruction overlap checks against each model's training data, or evaluate on a newly collected held-out set.
  2. [§5.6, Table 6] The user study protocol averages 4 expert ratings per dimension, but no inter-rater agreement statistic (e.g., Krippendorff's alpha, ICC, or per-dimension variance) is reported. The top-model differences in Table 6 are small on a 1–5 scale (3.72 vs. 3.70 vs. 3.61), and without confidence intervals or significance testing the claimed ranking is not statistically supported. The paper should report rating distributions, per-task standard deviations, and at least one agreement measure to justify treating averaged expert scores as reliable ground truth.
  3. [Data Availability, §5.5] The central artifact, CDD-IIE Bench, is not released: the Data Availability statement says it 'will be publicly released in the near future,' and the detailed rubrics are likewise promised but not provided. For a benchmark contribution, the 1,353 pairs with task labels, source attributions, per-dimension rubrics, and expert scores must be available for verification. Without release, the benchmark's claims of comprehensiveness and diagnosability cannot be checked, and the reported model scores are not reproducible.
  4. [§5.5.2, Table 3] There is an internal inconsistency in the benchmark's task count: the text and abstract state 21 tasks, while Table 3 lists 20 total tasks for CDD-IIE Bench. Additionally, the model naming is inconsistent across tables (e.g., 'Qwen-Edit14', 'Qwen-Edit22', 'Qwen-Edit6' in Tables 4, 5, and 6, with footnotes indicating they all refer to Qwen-Image-Edit-2509). These ambiguities should be corrected; more importantly, per-task sample counts and curation criteria for the internet-collected pairs are not specified, so the benchmark composition is not fully reproducible as described.
minor comments (4)
  1. [§5.5.2, §5.6] The manuscript refers to '12 professional annotators' in §5.5.2 and '12 visual experts' in §5.6; the relationship between the curating annotators and the evaluating experts should be clarified.
  2. [§3.2.1 vs §5.4.1] Two distinct works named 'AnyEdit' appear to be cited: [19] (ICML 2025, 'Edit any knowledge encoded in language models') is used in the data-construction sections, while [51] (CVPR 2024, 'Mastering Unified High-Quality Image Editing for Any Idea') is listed as a benchmark source. The manuscript should distinguish these clearly, since the benchmark composition in §5.5.2 depends on which AnyEdit is meant.
  3. [§5.7.1] There is a typo: 'This bottleneck represents a obstacle' should be 'an obstacle.'
  4. [Table 4/5 footnotes] The superscript footnotes mapping model names to tasks are easy to misread; consider using full model names in the table header or a separate column.

Circularity Check

1 steps flagged

CDD-IIE Bench's test pairs are curated from GEdit-Bench [22], the same reference as evaluated Step1X-Edit-v1.2, with no leakage analysis; the ranking is partially circular.

specific steps
  1. other [§5.5.2 Benchmark Composition; §5.7.1 Result Analysis; Tables 4–6]
    "These pairs are drawn from a diverse array of sources, including established datasets such as Gedit-Bench [22], I2EBench [124], ImgEdit-Bench [24], ComplexBench-Edit [135], CompBench [128], KRIS-Bench [137], AnyEdit [51], ByteMorph [138], Reason-Edit [48] and RefEdit-Bench [41] ... Step1X-Edit-v1.2 [22] shows the best performance in Implicit Reasoning."

    The CDD-IIE test pairs are curated from GEdit-Bench [22], and the evaluated model Step1X-Edit-v1.2 is the same reference [22]. Therefore, for that model, the benchmark is not an independent test set: the samples are drawn from the same line of work that produced the model. The paper reports no overlap/leakage analysis and has not released the benchmark, so the reported ranking and diagnostic conclusions (e.g., Step1X-Edit-v1.2 best at Implicit Reasoning) may reflect familiarity with the test distribution rather than editing ability. The claim that CDD-IIE 'can rigorously assess' model performance is thus partially circular for the empirical component.

full rationale

The survey portions—taxonomy, model review, metric discussion—are self-contained and do not reduce to their inputs. The circularity is confined to the proposed CDD-IIE Bench evaluation: it reuses existing benchmark data, including GEdit-Bench [22], while evaluating a model from that same work. Without overlap analysis or inter-rater reliability, the benchmark's 'rigorous' validity and the resulting model hierarchy are not independent. This is a partial circularity in the empirical claim, not in the survey's conceptual contributions.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 0 invented entities

The paper introduces no mathematical free parameters; its design choices (5 dimensions, 21 tasks, 3 rubric dimensions per task) are hand-chosen. The load-bearing assumptions are: expert scores are reliable, the benchmark is representative and uncontaminated, and CLIP-based filtering is a valid proxy for instruction adherence.

axioms (3)
  • domain assumption Human expert ratings on 1-5 rubrics, averaged over four raters per dimension, form a valid gold standard for edit quality.
    Used throughout §5.6 to produce scores in Tables 4-6; no inter-rater agreement, variance, or calibration against objective metrics is reported.
  • ad hoc to paper The 1,353 curated pairs are representative of the IIE task space and are unseen by evaluated models.
    The curation in §5.5.2 draws from existing benchmark datasets, including ones tied to evaluated models (GEdit-Bench/Step1X-Edit, AnyEdit), but no leakage analysis is provided.
  • domain assumption CLIP-based text-image similarity and related filtering metrics reliably indicate instruction adherence and edit quality.
    The data construction pipelines in §3 rely heavily on CLIP similarity thresholds to filter training pairs; this is a known approximate proxy and the paper does not validate it for the newly proposed benchmark.

pith-pipeline@v1.3.0-alltime-deepseek · 39449 in / 11154 out tokens · 126207 ms · 2026-08-01T01:48:28.549061+00:00 · methodology

0 comments
read the original abstract

Instruction-based Image Editing (IIE) aims to transform a given image into a new one based on textual instructions. Advances in Large Language Models (LLMs) and Vision-Language Models (VLMs) have accelerated progress toward practical ``one-sentence image editing" systems. This survey presents a systematic taxonomy and comprehensive review of IIE research, structured around five core dimensions: (1) task definition and hierarchical categorization of editing operations, (2) methodologies for training data construction, (3) architectural evolution from GAN-based to diffusion and autoregressive paradigms, (4) standardized evaluation metrics and benchmark development, and (5) introduction of commercial solutions. Our analysis shows critical technological milestones across model generations. We further propose a Comprehensive, in-Depth, and Diagnostic benchmark for IIE task (CDD-IIE Bench), which can rigorously assess the multiple aspects of model performance. Through empirical comparisons of open-source solutions, we highlight their respective capabilities and limitations. Finally, we discuss future research directions to advance the field.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

148 extracted references · 27 linked inside Pith

  1. [1]

    Goodfellow, J

    I.J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, Y. Bengio,Genera- tive adversarial nets, inAdvances in Neural Information Processing Systems(Curran Associates, Montreal, Canada, 2014), pp. 2672–2680

    2014

  2. [2]

    Baykal, A.B

    A.C. Baykal, A.B. Anees, D. Ceylan, E. Erdem, A. Erdem, D. Yuret, Clip-guided stylegan inversion for text-driven real image editing. ACM Trans. Graph.42(5), 1–18 (2023)

    2023

  3. [3]

    Karras, S

    T. Karras, S. Laine, T. Aila,A Style- Based Generator Architecture for Genera- tive Adversarial Networks, inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(IEEE, Long Beach, USA, 2019), pp. 4401–4410

    2019

  4. [4]

    Pathak, P

    D. Pathak, P. Krahenbuhl, J. Donahue, T. Darrell, A.A. Efros,Context encoders: Feature learning by inpainting, inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (IEEE, Las Vegas, USA, 2016), pp. 2536– 2544

    2016

  5. [5]

    Ledig, L

    C. Ledig, L. Theis, F. Husz´ ar, J. Caballero, A. Cunningham, A. Acosta, A. Aitken, A. Tejani, J. Totz, Z. Wang, et al.,Photo- realistic single image super-resolution using a generative adversarial network, inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (IEEE, Honolulu, USA, 2017), pp. 4681– 4690

    2017

  6. [6]

    Patashnik, Z

    O. Patashnik, Z. Wu, E. Shechtman, D. Cohen-Or, D. Lischinski,Styleclip: Text- driven manipulation of stylegan imagery, in Proceedings of the IEEE/CVF International Conference on Computer Vision(IEEE, Virtual, 2021), pp. 2085–2094

    2021

  7. [7]

    Hertz, R

    A. Hertz, R. Mokady, J. Tenenbaum, K. Aberman, Y. Pritch, D. Cohen-Or, Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022)

    Pith/arXiv arXiv 2022

  8. [8]

    Couairon, J

    G. Couairon, J. Verbeek, H. Schwenk, M. Cord,Diffedit: Diffusion-based seman- tic image editing with mask guidance, in Proceedings of the International Confer- ence on Learning Representations(OpenRe- view.net, Virtual, 2022), pp. 1–20

    2022

  9. [9]

    Lugmayr, M

    A. Lugmayr, M. Danelljan, A. Romero, F. Yu, R. Timofte, L. Van Gool,Repaint: Inpainting using denoising diffusion prob- abilistic models, inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(IEEE, New Orleans, USA, 2022), pp. 11461–11471 34

    2022

  10. [10]

    Avrahami, D

    O. Avrahami, D. Lischinski, O. Fried, Blended diffusion for text-driven edit- ing of natural images, inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(IEEE, New Orleans, USA, 2022), pp. 18208–18218

    2022

  11. [11]

    L. Yu, B. Shi, R. Pasunuru, B. Muller, O. Golovneva, T. Wang, A. Babu, B. Tang, B. Karrer, S. Sheynin, Scaling autore- gressive multi-modal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591 (2023)

    Pith/arXiv arXiv 2023

  12. [12]

    G.H. Wang, S. Zhao, X. Zhang, L. Cao, P. Zhan, L. Duan, S. Lu, M. Fu, X. Chen, J. Zhao, Ovis-u1 technical report. arXiv preprint arXiv:2506.23044 (2025)

    Pith/arXiv arXiv 2025

  13. [13]

    C. Wu, P. Zheng, R. Yan, S. Xiao, X. Luo, Y. Wang, W. Li, X. Jiang, Y. Liu, J. Zhou, Omnigen2: Exploration to advanced multimodal generation. arXiv preprint arXiv:2506.18871 (2025)

    Pith/arXiv arXiv 2025

  14. [14]

    J. Lu, C. Clark, S. Lee, Z. Zhang, S. Khosla, R. Marten, D. Hoiem, A. Kembhavi, Unified-io 2: Scaling autoregressive multi- modal models with vision language audio and action, inProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition(IEEE, Seattle, USA, 2024), pp. 26439–26455

    2024

  15. [15]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, B. Ommer,High-Resolution Image Synthesis with Latent Diffusion Models, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion(IEEE, New Orleans, USA, 2022), pp. 10684–10695

    2022

  16. [16]

    Zhang, P

    R. Zhang, P. Isola, A.A. Efros, E. Shecht- man, O. Wang,The unreasonable effec- tiveness of deep features as a perceptual metric, inProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition(IEEE, Long Beach, USA, 2018), pp. 586–595

    2018

  17. [17]

    Johnson, A

    J. Johnson, A. Alahi, L. Fei-Fei,Perceptual losses for real-time style transfer and super- resolution, inEuropean Conference on Com- puter Vision(Springer, Amsterdam, The Netherlands, 2016), pp. 694–711

    2016

  18. [18]

    Sheynin, A

    S. Sheynin, A. Polyak, U. Singer, Y. Kirstain, A. Zohar, O. Ashual, D. Parikh, Y. Taigman,Emu edit: Precise image edit- ing via recognition and generation tasks, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition(IEEE, Seattle, USA, 2024), pp. 8871–8879

    2024

  19. [19]

    Jiang, J

    H. Jiang, J. Fang, N. Zhang, G. Ma, M. Wan, X. Wang, X. He, T.s. Chua, Anyedit: Edit any knowledge encoded in lan- guage models, inProceedings of the Inter- national Conference on Machine Learn- ing(ACM, Vancouver, Canada, 2025), pp. 27510–27533

    2025

  20. [20]

    S. Liu, Z. Zeng, T. Ren, F. Li, H. Zhang, J. Yang, Q. Jiang, C. Li, J. Yang, H. Su, et al.,Grounding dino: Marrying dino with grounded pre-training for open-set object detection, inEuropean Conference on Com- puter Vision(Springer, Milan, Italy, 2024), pp. 38–55

    2024

  21. [21]

    Kirillov, E

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. White- head, A.C. Berg, W.Y. Lo, et al.,Seg- ment anything, inProceedings of the IEEE/CVF International Conference on Computer Vision(IEEE, Paris, France, 2023), pp. 4015–4026

    2023

  22. [22]

    S. Liu, Y. Han, P. Xing, F. Yin, R. Wang, W. Cheng, J. Liao, Y. Wang, H. Fu, C. Han, Step1x-edit: A practical framework for general image editing. arXiv preprint arXiv:2504.17761 (2025)

    Pith/arXiv arXiv 2025

  23. [23]

    arXiv preprint arXiv:2502.13923 (2025)

    Qwen Team, Qwen2.5-vl technical report. arXiv preprint arXiv:2502.13923 (2025)

    Pith/arXiv arXiv 2025

  24. [24]

    Y. Ye, X. He, Z. Li, B. Lin, S. Yuan, Z. Yan, B. Hou, L. Yuan,Imgedit: A unified image editing dataset and benchmark, inThe Annual Conference on Neural Information 35 Processing Systems Datasets and Bench- marks Track(Curran Associates, San Diego, USA, 2025), pp. 1–20

    2025

  25. [25]

    Zhang, X

    Y. Zhang, X. Huang, J. Ma, Z. Li, Z. Luo, Y. Xie, Y. Qin, T. Luo, Y. Li, S. Liu, et al., Recognize anything: A strong image tagging model, inProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition(IEEE, Seattle, USA, 2024), pp. 1724–1732

    2024

  26. [26]

    N. Ravi, V. Gabeur, Y.T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. R¨ adle, C. Rolland, L. Gustafson, et al.,Sam 2: Segment anything in images and videos, inProceedings of the International Confer- ence on Learning Representations(OpenRe- view.net, Vienna, Austria, 2024), pp. 1–20

    2024

  27. [27]

    Zhang, A

    L. Zhang, A. Rao, M. Agrawala,Adding Conditional Control to Text-to-Image Dif- fusion Models, inProceedings of the IEEE/CVF International Conference on Computer Vision(IEEE, Paris, France, 2023), pp. 3836–3847

    2023

  28. [28]

    E.J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chen, et al., Lora: Low-rank adaptation of large lan- guage models., inProceedings of the Interna- tional Conference on Learning Representa- tions(OpenReview.net, Virtual, 2022), pp. 1–17

    2022

  29. [29]

    Q. Wang, Y. Shi, J. Ou, R. Chen, K. Lin, J. Wang, B. Jiang, H. Yang, M. Zheng, X. Tao, et al.,Koala-36m: A large-scale video dataset improving consistency between fine-grained conditions and video content, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition(IEEE, Nashville, USA, 2025), pp. 8428–8437

    2025

  30. [30]

    Zheng, D

    P. Zheng, D. Gao, D.P. Fan, L. Liu, J. Laak- sonen, W. Ouyang, N. Sebe, Bilateral refer- ence for high-resolution dichotomous image segmentation. CAAI Artif. Intell. Res.3, 1–12 (2024)

    2024

  31. [31]

    Z. Teed, J. Deng,Raft: Recurrent all-pairs field transforms for optical flow, inEuropean Conference on Computer Vision(Springer, Glasgow, UK, 2020), pp. 402–419

    2020

  32. [32]

    arXiv preprint arXiv:2410.21276 (2024)

    OpenAI, Gpt-4o system card. arXiv preprint arXiv:2410.21276 (2024)

    Pith/arXiv arXiv 2024

  33. [33]

    Brooks, A

    T. Brooks, A. Holynski, A.A. Efros, Instructpix2pix: Learning to follow image editing instructions, inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(IEEE, Vancouver, Canada, 2023), pp. 18392–18402

    2023

  34. [34]

    Y. Du, C. Li, R. Guo, X. Yin, W. Liu, J. Zhou, Y. Bai, Z. Yu, Y. Yang, Q. Dang, et al., Pp-ocr: a practical ultra lightweight ocr system. arXiv preprint arXiv:2009.09941 (2009)

    Pith/arXiv arXiv 2009

  35. [35]

    Martin, C

    D. Martin, C. Fowlkes, D. Tal, J. Malik, A database of human segmented natu- ral images and its application to evaluat- ing segmentation algorithms and measur- ing ecological statistics, inProceedings of the IEEE/CVF International Conference on Computer Vision(IEEE, Vancouver, Canada, 2001), pp. 416–423

    2001

  36. [36]

    Nah, T.H

    S. Nah, T.H. Kim, K.M. Lee,Deep multi-scale convolutional neural network for dynamic scene deblurring, inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(IEEE, Honolulu, USA, 2017), pp. 3883–3891

    2017

  37. [37]

    Cheng, X

    D. Cheng, X. Li, W.H. Li, C. Lu, F. Li, H. Zhao, W.S. Zheng,Large-Scale Vis- ible Watermark Detection and Removal with Deep Convolutional Networks, inPro- ceedings of the Chinese Conference on Pattern Recognition and Computer Vision (Springer, Guangzhou, China, 2018), pp. 27–40

    2018

  38. [38]

    J.Y. Rim, G. Lee, J. Won, S. Cho, Real-World Blur Dataset for Learning and Benchmarking Deblurring Algorithms, in European Conference on Computer Vision (Springer, Glasgow, UK, 2020), pp. 184–201 36

    2020

  39. [39]

    Abdelhamed, S

    A. Abdelhamed, S. Lin, M.S. Brown, A High-Quality Denoising Dataset for Smartphone Cameras, inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(IEEE, Salt Lake City, USA, 2018), pp. 1692–1700

    2018

  40. [40]

    Y. Liu, Z. Zhu, X. Bai,Wdnet: Watermark- decomposition network for visible watermark removal, inProceedings of the IEEE/CVF Winter Conference on Applications of Com- puter Vision(IEEE, Virtual, 2021), pp. 3685–3693

    2021

  41. [41]

    Pathiraja, M

    B. Pathiraja, M. Patel, S. Singh, Y. Yang, C. Baral,RefEdit: A Benchmark and Method for Improving Instruction-based Image Editing Model on Referring Expres- sions, inProceedings of the IEEE/CVF International Conference on Computer Vision(IEEE, Marrakech, Morocco, 2025), pp. 15646–15656

    2025

  42. [42]

    Karras, S

    T. Karras, S. Laine, M. Aittala, J. Hellsten, J. Lehtinen, T. Aila,Analyzing and Improv- ing the Image Quality of StyleGAN, inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (IEEE, Seattle, USA, 2020), pp. 8107–8116

    2020

  43. [43]

    T. Wei, D. Chen, W. Zhou, J. Liao, Z. Tan, L. Yuan, W. Zhang, N. Yu,Hairclip: Design your hair by text and reference image, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion(IEEE, New Orleans, USA, 2022), pp. 18072–18081

    2022

  44. [44]

    Kwon, J.C

    G. Kwon, J.C. Ye,CLIPstyler: Image Style Transfer with a Single Text Condition, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion(IEEE, New Orleans, USA, 2022), pp. 18041–18050

    2022

  45. [45]

    W. Xia, Y. Yang, J.H. Xue, B. Wu,Tedi- gan: Text-guided diverse face image gener- ation and manipulation, inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(IEEE, Vir- tual, 2021), pp. 2256–2265

    2021

  46. [46]

    Jiang, Z

    Y. Jiang, Z. Huang, X. Pan, C.C. Loy, Z. Liu,Talk-to-edit: Fine-grained facial editing via dialog, inProceedings of the IEEE/CVF International Conference on Computer Vision(IEEE, Virtual, 2021), pp. 13799–13808

    2021

  47. [47]

    Zhang, L

    K. Zhang, L. Mo, W. Chen, H. Sun, Y. Su,Magicbrush: A manually annotated dataset for instruction-guided image editing, inAdvances in Neural Information Pro- cessing Systems(Curran Associates, New Orleans, USA, 2023), pp. 31428–31449

    2023

  48. [48]

    Huang, L

    Y. Huang, L. Xie, X. Wang, Z. Yuan, X. Cun, Y. Ge, J. Zhou, C. Dong, R. Huang, R. Zhang, Y. Shan,SmartEdit: Exploring Complex Instruction-Based Image Editing with Multimodal Large Language Models, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion(IEEE, Seattle, USA, 2024), pp. 8362– 8371

    2024

  49. [49]

    L. Yang, Z. Yu, C. Meng, M. Xu, S. Ermon, B. Cui,Mastering Text-to-Image Diffu- sion: Recaptioning, Planning, and Generat- ing with Multimodal LLMs, inProceedings of the International Conference on Learning Representations(OpenReview.net, Vienna, Austria, 2024), pp. 1–20

    2024

  50. [50]

    X. Ju, X. Liu, X. Wang, Y. Bian, Y. Shan, Q. Xu,Brushnet: A plug-and-play image inpainting model with decomposed dual- branch diffusion, inEuropean Conference on Computer Vision(Springer, Milan, Italy, 2024), pp. 150–168

    2024

  51. [51]

    Q. Yu, W. Chow, Z. Yue, K. Pan, Y. Wu, X. Wan, J. Li, S. Tang, H. Zhang, Y. Zhuang,AnyEdit: Mastering Unified High-Quality Image Editing for Any Idea, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion(IEEE, Seattle, USA, 2024), pp. 26125– 26135

    2024

  52. [52]

    Zhao, X.S

    H. Zhao, X.S. Ma, L. Chen, S. Si, R. Wu, K. An, P. Yu, M. Zhang, Q. Li, B. Chang, UltraEdit: Instruction-based Fine-Grained 37 Image Editing at Scale, inAdvances in Neu- ral Information Processing Systems(Curran Associates, Vancouver, Canada, 2024), pp. 3058–3093

    2024

  53. [53]

    Z. Geng, B. Yang, T. Hang, C. Li, S. Gu, T. Zhang, J. Bao, Z. Zhang, H. Li, H. Hu,Instructdiffusion: A generalist mod- eling interface for vision tasks, inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (IEEE, Seattle, USA, 2024), pp. 12709– 12720

    2024

  54. [54]

    Y. Jin, P. Ling, X. Dong, P. Zhang, J. Wang, D. Lin, Reasonpix2pix: instruction reason- ing dataset for advanced image editing. arXiv preprint arXiv:2405.11190 (2024)

    Pith/arXiv arXiv 2024

  55. [55]

    Shagidanov, H

    A. Shagidanov, H. Poghosyan, X. Gong, Z. Wang, S. Navasardyan, H. Shi,Grounded- Instruct-Pix2Pix: Improving Instruction Based Image Editing with Automatic Target Grounding, inIEEE International Con- ference on Acoustics, Speech and Signal Processing(IEEE, Seoul, Korea, 2024), pp. 6585–6589

    2024

  56. [56]

    M. Li, X. Gu, F. Chen, X. Xing, L. Wen, C. Chen, S. Zhu,Superedit: Rectifying and facilitating supervision for instruction- based image editing, inProceedings of the IEEE/CVF International Conference on Computer Vision(IEEE, Marrakech, Morocco, 2025), pp. 19206–19215

    2025

  57. [57]

    Kawar, S

    B. Kawar, S. Zada, O. Lang, O. Tov, H. Chang, T. Dekel, I. Mosseri, M. Irani, Imagic: Text-based real image editing with diffusion models, inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(IEEE, Vancouver, Canada, 2023), pp. 6007–6017

    2023

  58. [58]

    Mokady, A

    R. Mokady, A. Hertz, K. Aberman, Y. Pritch, D. Cohen-Or,Null-text inver- sion for editing real images using guided diffusion models, inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(IEEE, Vancouver, Canada, 2023), pp. 6038–6047

    2023

  59. [59]

    Valevski, M

    D. Valevski, M. Kalman, E. Molad, E. Segalis, Y. Matias, Y. Leviathan, Uni- tune: Text-driven image editing by fine tun- ing a diffusion model on a single image. ACM Trans. Graph.42(4), 1–10 (2023)

    2023

  60. [60]

    Avrahami, K

    O. Avrahami, K. Aberman, O. Fried, D. Cohen-Or, D. Lischinski,Break-a-scene: Extracting multiple concepts from a single image, inProceedings of the SIGGRAPH Asia Conference Papers(ACM, Sydney, Australia, 2023), pp. 1–12

    2023

  61. [61]

    Jiang, X

    R. Jiang, X. Fu, G. Zheng, T. Li, T. Yao, X. Li,Energy-guided optimization for per- sonalized image editing with pretrained text- to-image diffusion models, inProceedings of the AAAI Conference on Artificial Intelli- gence, vol. 39 (AAAI, Philadelphia, USA, 2025), pp. 4048–4056

    2025

  62. [62]

    W. Dong, S. Xue, X. Duan, S. Han,Prompt Tuning Inversion for Text-Driven Image Editing Using Diffusion Models, inProceed- ings of the IEEE/CVF International Con- ference on Computer Vision(IEEE, Paris, France, 2023), pp. 7396–7406

    2023

  63. [63]

    S. Li, J. van de Weijer, T. Hu, F.S. Khan, Q. Hou, Y. Wang, J. Yang, Stylediffusion: Prompt-embedding inversion for text-based editing. arXiv preprint arXiv:2303.15649 (2023)

    Pith/arXiv arXiv 2023

  64. [64]

    Q. Wang, B. Zhang, M. Birsak, P. Wonka, Instructedit: Improving automatic masks for diffusion-based image editing with user instructions. arXiv preprint arXiv:2305.18047 (2023)

    Pith/arXiv arXiv 2023

  65. [65]

    Parmar, K

    G. Parmar, K. Kumar Singh, R. Zhang, Y. Li, J. Lu, J.Y. Zhu,Zero-shot image- to-image translation, inACM SIGGRAPH 2023 Conference Proceedings(ACM, New York, USA, 2023), pp. 1–11

    2023

  66. [66]

    Y. Qiao, F. Wang, J. Su, Y. Zhang, Y. Yu, S. Wu, G.J. Qi,Baret: Balanced atten- tion based real image editing driven by target-text inversion, inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38 (AAAI, Vancouver, Canada, 2024), 38 pp. 4560–4568

    2024

  67. [67]

    C. Meng, Y. He, Y. Song, J. Song, J. Wu, J. Zhu, S. Ermon,SDEdit: Guided Image Synthesis and Editing with Stochas- tic Differential Equations, inProceedings of the International Conference on Learning Representations(OpenReview.net, Virtual, 2022), pp. 1–20

    2022

  68. [68]

    Tumanyan, M

    N. Tumanyan, M. Geyer, S. Bagon, T. Dekel,Plug-and-play diffusion features for text-driven image-to-image translation, inProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition(IEEE, Vancouver, Canada, 2023), pp. 1921–1930

    2023

  69. [69]

    Titov, M

    V. Titov, M. Khalmatova, A. Ivanova, D. Vetrov, A. Alanov,Guide-and-rescale: Self-guidance mechanism for effective tuning-free real image editing, inEuropean Conference on Computer Vision(Springer, Milan, Italy, 2024), pp. 235–251

    2024

  70. [70]

    S. Li, B. Zeng, Y. Feng, S. Gao, X. Liu, J. Liu, L. Li, X. Tang, Y. Hu, J. Liu,Zone: Zero-shot instruction-guided local editing, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion(IEEE, Seattle, USA, 2024), pp. 6254– 6263

    2024

  71. [71]

    Y. Li, Y. Bian, X. Ju, Z. Zhang, Y. Shan, Y. Zou, Q. Xu, Brushedit: All-in-one image inpainting and editing. arXiv preprint arXiv:2412.10316 (2024)

    Pith/arXiv arXiv 2024

  72. [72]

    Simsar, A

    E. Simsar, A. Tonioni, Y. Xian, T. Hof- mann, F. Tombari,Uip2p: Unsupervised instruction-based image editing via cycle edit consistency, inProceedings of the IEEE/CVF International Conference on Computer Vision(IEEE, Marrakech, Morocco, 2025), pp. 18895–18905

    2025

  73. [73]

    Brack, F

    M. Brack, F. Friedrich, K. Kornmeier, L. Tsaban, P. Schramowski, K. Kersting, A. Passos,Ledits++: Limitless image edit- ing using text-to-image models, inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (IEEE, Seattle, USA, 2024), pp. 8861–8870

    2024

  74. [74]

    M. Cao, X. Wang, Z. Qi, Y. Shan, X. Qie, Y. Zheng,Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing, inProceedings of the IEEE/CVF International Conference on Computer Vision(IEEE, Paris, France, 2023), pp. 22560–22570

    2023

  75. [75]

    Hertz, K

    A. Hertz, K. Aberman, D. Cohen-Or, Delta denoising score, inProceedings of the IEEE/CVF International Conference on Computer Vision(IEEE, Paris, France, 2023), pp. 2328–2337

    2023

  76. [76]

    X. Ju, A. Zeng, Y. Bian, S. Liu, Q. Xu,PnP Inversion: Boosting Diffusion-based Editing with 3 Lines of Code, inProceedings of the International Conference on Learning Representations(OpenReview.net, Vienna, Austria, 2024), pp. 1–20

    2024

  77. [77]

    M. Arar, A. Voynov, S. Fruchter, D. Cohen- Or,Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models, inProceedings of the IEEE/CVF Winter Conference on Applica- tions of Computer Vision(IEEE, Tucson, USA, 2025), pp. 2063–2072

    2025

  78. [78]

    Wallace, A

    B. Wallace, A. Gokul, N. Naik,EDICT: Exact Diffusion Inversion via Coupled Transformations, inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(IEEE, Vancouver, Canada, 2023), pp. 22532–22541

    2023

  79. [79]

    L. Han, S. Wen, Q. Chen, Z. Zhang, K. Song, M. Ren, R. Gao, A. Stathopoulos, X. He, Y. Chen, D. Liu, Q. Zhangli, J. Jiang, Z. Xia, A. Srivastava, D.N. Metaxas,Prox- Edit: Improving Tuning-Free Real Image Editing with Proximal Guidance, inProceed- ings of the IEEE/CVF Winter Conference on Applications of Computer Vision(IEEE, Waikoloa, USA, 2024), pp. 4279–4289

    2024

  80. [80]

    Huberman-Spiegelglas, V

    I. Huberman-Spiegelglas, V. Kulikov, T. Michaeli,An Edit Friendly DDPM Noise Space: Inversion and Manipulations, in Proceedings of the IEEE/CVF Conference 39 on Computer Vision and Pattern Recog- nition(IEEE, Seattle, USA, 2024), pp. 12469–12478

    2024

Showing first 80 references.