Pith. sign in

REVIEW 3 major objections 4 minor 46 references

This paper argues that localizing AI-generated forgeries works better when the final mask is built by progressive, evidence-guided updates to an initial state rather than a single one-shot prediction, with the largest gains on forgeries nev

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

2026-08-03 12:33 UTC pith:WJY65IML

load-bearing objection Sensible progressive-localization idea, clear writing, thorough ablations—but the headline OOD gain on AutoSplice is undercut because key hyperparameters were tuned on that same test set. the 3 major comments →

arxiv 2607.29156 v1 pith:WJY65IML submitted 2026-07-31 cs.CV

Progressive Decision-Making for Localizing Open-Ended AI-Generated Image Forgeries

Jingyi Hou , Xiaoxia Chen , Leyu Zhou , Zhichuang Wang , Zhijie Liu This is my paper
classification cs.CV
keywords image forgery localizationAI-generated image forgeryprogressive decision-makingsequential state updatinguncertainty estimationstate-space modelmesoscopic forensic tracesout-of-distribution generalization
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.

AI-generated image forgeries leave forensic traces that are real but weak and unevenly reliable across regions. The paper's central claim is that making the final localization mask through a single one-shot prediction wastes that evidence, because one fixed rule cannot handle regions that need different levels of caution. Instead, the authors reformulate final localization as a progressive decision-updating process: start with an initial mask, then update it a small number of times, guided by compacted evidence, an uncertainty map, and boundary cues, while preserving confident areas and revising ambiguous ones. The payoff, if the claim holds, is better masks on both conventional and AI-generated forgeries, with larger relative gains on AI-generated forgeries that were never shown in training. A sympathetic reader would take the contribution to be a new formulation of the final localization step, not a new forensic feature.

Core claim

The central claim is that the final forgery mask should be an evolving decision state, not a one-shot output. The method starts with an initial mask and updates it over a few steps with an Evidence-Guided Mamba operator, guided by compact decision evidence, a Shannon-entropy uncertainty map, and a boundary prior. A region-aware constraint preserves confident foreground/background pixels and revises only ambiguous ones. Trained solely on conventional manipulations, it raises average F1 on two AI-generated benchmarks to 0.4105 versus 0.3717 for the best competitor (about 10.4 percent relative), while conventional F1 improves about 5.2 percent. Adding AI-generated training data improves both be

What carries the argument

Two components carry the argument. A lightweight Decision Evidence Projector (a 1x1 convolution) compresses mesoscopic forensic features into compact decision evidence, so the update operates on decision-relevant responses rather than raw features. The Evidence-Guided Mamba (EG-Mamba) is a selective state-space sequence model used as the state-transition operator: it takes the concatenation of the evidence, current state, Shannon-entropy uncertainty map, and boundary-aware prior, and outputs a revised localization state. The region-aware constraint (three soft masks with thresholds tau_low and tau_high) then decides which pixels to preserve as confident foreground/background and which to act

Load-bearing premise

The load-bearing premise is that the AI-generated benchmark used for evaluation is genuinely unseen for the reported configuration; the paper selects thresholds, the number of update steps, and a threshold shift using that benchmark's performance, so if those choices count as test-set tuning, the claimed out-of-distribution gain of about 10 percent is partly circular.

What would settle it

Re-run the protocol with hyperparameters fixed using only conventional validation data (no AutoSplice in any selection step) and no inference-time threshold shift; if the average F1 on AI-generated benchmarks falls to the 0.3717 level of the best compared method, the progressive-updating generalization claim is refuted for this protocol. Alternatively, retrain the same model with T=0 (no progressive updates) and check whether the remaining gap to the full model disappears.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If correct, the method implies that a fixed post-hoc refinement of a one-shot mask is not the same as a decision-guided update; the gain comes from conditioning revision on uncertainty and boundary evidence at each step.
  • Progressive updating appears to be most useful exactly where evidence is weak and heterogeneous, so AI-generated forgeries benefit more than conventional ones when training contains only conventional manipulations.
  • The benefit saturates: two update steps perform best, and more steps degrade performance, suggesting that the localization state settles quickly and that supervision of intermediate states is doing the work.
  • The inference-time threshold shift shows the initial states are conservative; masking confident regions during updates is essential, since unconstrained iterative refinement hurts.
  • The robustness experiments indicate the updating process helps under Gaussian noise, blur, and JPEG compression, particularly on out-of-distribution AI-generated forgeries.

Where Pith is reading between the lines

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

  • An untested extension follows directly from the paper's logic: the same progressive decision-updating stage should transfer to other forensic backbones, not just the mesoscopic-trace one used here, and would be a stronger demonstration of the formulation's independence from the evidence extractor.
  • Because the uncertainty map is computed once from the initial state and held fixed, the paper implicitly commits to the initial state carrying all ambiguity information; a reader could test whether recomputing uncertainty from intermediate states helps or destabilizes the update.
  • The claim that progressive updating generalizes to 'open-ended' forgeries suggests a testable scaling prediction: performance on a new generative model should degrade more gracefully with this method than with one-shot baselines as the manipulation distribution shifts further from training.
  • The threshold parameters and inference-time shift are selected using the AI benchmark's validation performance; an honest deployment would fix them on conventional data only, and the reported 10 percent out-of-distribution gain is the number to re-examine under that constraint.

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

3 major / 4 minor

Summary. The paper proposes a progressive decision-updating formulation for image forgery localization, replacing one-shot mask prediction with an iterative update process. A Decision Evidence Projector converts mesoscopic features into compact evidence, and an Evidence-Guided Mamba (EG-Mamba) updates the localization state using the current map, a fixed Shannon-entropy uncertainty cue, and a boundary-aware prior, under a region-aware constraint that preserves confident foreground/background while revising uncertain pixels. The method is evaluated on conventional benchmarks (NIST16, CASIAv1) and AI-generated benchmarks (AutoSplice, SAGI-D-9K) under Protocol-CAT training and a second protocol that introduces SAGI-D-9K training data. The central claim is that progressive updating gives larger gains on unseen AI-generated forgeries than one-shot prediction, with the headline numbers in Tables I and II supporting this claim.

Significance. If the empirical generalization claim held, the paper would offer a useful and fresh perspective: it moves the design focus from feature extraction to the final decision stage and provides systematic ablations showing that each component (evidence projector, uncertainty cue, boundary prior, region-aware constraint, deep supervision) contributes to the reported performance. The method is clearly described, and the planned code release is a strength. However, the key out-of-distribution claim is currently compromised because the thresholds, update-step count, and inference-time threshold calibration are selected on AutoSplice, which is itself one of the test benchmarks described as 'unseen'. The methodological idea remains plausible, but the headline quantitative result needs a clean evaluation protocol before it can be judged.

major comments (3)
  1. [Sec. IV-F2/F3; Sec. IV-C; Tables I-II] The hyperparameters tau_low=0.3, tau_high=0.8, and T=2 are selected by scanning performance on AutoSplice under Protocol II (Figs. 3-4), and Sec. IV-C fixes these values 'according to the validation analysis in Sec. IV-F3'. In addition, the inference-time threshold shift of -0.2 (Sec. III-D2) is not validated on any held-out set. AutoSplice is then reported as an 'unseen' test benchmark in Tables I and II and described as out-of-distribution in Sec. IV-E. No validation split of AutoSplice is mentioned. The headline claim that the method 'brings larger gains on unseen AI-generated forgeries' (Abstract and Sec. V) is therefore at risk of selection circularity: the reported AutoSplice gains may partly reflect hyperparameter fitting to that test set. This is a load-bearing issue for the central generalization claim. The authors should either hold out a validation split of AutoSplice and veri
  2. [Table III; Figs. 3-4] The component ablations and parameter analyses are all performed on AutoSplice under Protocol II. Because AutoSplice is also the benchmark used to justify the design choices, the ablations do not provide an independent check of the design assumptions on an out-of-distribution dataset. At least one ablation on a held-out source (e.g., NIST16 or CASIAv1) would clarify whether the component gains are specific to AutoSplice or generalize; this is particularly important given that the paper's central claim is about generalization.
  3. [Table I; Sec. IV-D] Co-Transformer results are reported from the original paper without official code, while other methods are reproduced under the same evaluation protocol. This can introduce a mismatch in training and evaluation conditions for a method that contributes to the conventional benchmark comparison. Please confirm explicitly that Co-Transformer's published numbers are obtained under the same Protocol-CAT training and evaluation setting, or provide a reproducibility matrix. Also clarify how the SAGI-D-9K subset was sampled and whether the split is reproducible.
minor comments (4)
  1. [Sec. IV-F3] Typo: 'BCE nad Dice' should read 'BCE and Dice'.
  2. [Table IV] The column heading notation '3 / 100', '7 / 90', etc. is ambiguous because the meaning of the two numbers differs between perturbation types. Please separate noise/blur and JPEG into distinct sub-tables or use explicit headers such as 'sigma / quality'.
  3. [Sec. IV-G; Fig. 7] The visualization discussion states that the initial state P0 already provides more complete localization responses, but this is not quantified. Adding the P0 performance to the ablation table would make the claim verifiable.
  4. [Table IV] The robustness comparison only includes Mesorch as the baseline. Adding one or two of the stronger compared methods (e.g., PSCC-Net or IML-ViT) would make the robustness claim more convincing.

Circularity Check

1 steps flagged

AutoSplice is used both to select key hyperparameters (τ_low, τ_high, T) and as an 'unseen' AIGC test benchmark, making the headline OOD generalization claim partially selection-circular; the architecture itself retains independent empirical support.

specific steps
  1. fitted input called prediction [Sec. IV-C and Sec. IV-F1/F3 (Figs. 3-4); AutoSplice rows in Tables I-II]
    "We conduct ablation studies on AutoSplice under Protocol II ... As shown in Fig. 3, the best performance is obtained when τ_low = 0.3 and τ_high = 0.8, with an F1-score of 0.6412 and a Permute F1-score of 0.7607. ... Based on this analysis, we use T=2 during inference in our experiments. ... Since AutoSplice is not used for training, it still serves as an out-of-distribution AI-generated manipulation benchmark."

    The paper selects τ_low=0.3, τ_high=0.8, and T=2 by maximizing F1/Permute-F1 on AutoSplice, then reports AutoSplice in Tables I and II as an out-of-distribution/'unseen' AIGC benchmark and uses those same scores to support the claim of larger gains on unseen AI-generated forgeries. Because the reported AutoSplice numbers are the optimization surface for the model-selection procedure rather than a held-out test, the AutoSplice-specific generalization claim is partly fitted. This does not make the architecture definitionally circular, but it means the headline 'unseen' advantage on AutoSplice is not an independent evaluation; the SAGI-D-9K and conventional-benchmark results provide separate, cleaner evidence.

full rationale

Most of this paper is an empirical architecture paper rather than a formal derivation, so the core comparison against one-shot baselines is not circular: the proposed module is trained and evaluated on external benchmarks, and the ablations isolate the contributions of the evidence projector, uncertainty cue, boundary prior, and region-aware constraint. I found no load-bearing self-citation chain and no equation-level equivalence between inputs and outputs. The one significant issue is evaluation-protocol circularity: the paper tunes key hyperparameters (τ_low=0.3, τ_high=0.8, T=2, and apparently λ_deep=1.0) by scanning performance on AutoSplice in Sec. IV-F, then describes AutoSplice as an out-of-distribution AI-generated benchmark in the main comparisons. Thus the AutoSplice portion of the 'larger gains on unseen AI-generated forgeries' claim is at least partially selection-circular, since no held-out validation split is described and the test set was used to choose the reported configuration. However, the central claim still has independent content: the method also improves on SAGI-D-9K under Protocol I (0.2622 vs. 0.2103 for the best competitor), and the conventional-benchmark gains are not affected by the AutoSplice tuning. A moderate score of 4 reflects this partial circularity while acknowledging that the architecture is not itself derived from its own outputs.

Axiom & Free-Parameter Ledger

8 free parameters · 5 axioms · 0 invented entities

No new physical or conceptual entities are introduced. 'Decision evidence' is a projected feature map and EG-Mamba is an architecture module, not a newly postulated entity. The main burden is carried by hand-selected hyperparameters and by the assumption that mesoscopic features plus progressive updating can compensate for weak AIGC traces.

free parameters (8)
  • tau_low / tau_high = 0.3 / 0.8
    Confidence thresholds in the region-aware constraint; chosen by grid search in Fig. 3 on AutoSplice under Protocol II; directly control how much of the mask is preserved versus revised.
  • inference threshold shift = -0.2
    Applied to tau_high and tau_low at test time; motivated by conservative initial responses but appears manually tuned rather than derived; affects all final results.
  • alpha (boundary prior scaling) = not reported
    Scaling factor in B = exp(-alpha*G) (Eq. 7); controls the strength of the boundary propagation penalty; value not specified in the manuscript.
  • update steps T = 2 at inference
    Selected from validation analysis in Fig. 4 on AutoSplice; used in final reported results; training is unrolled with T=2.
  • beta (mask sharpness) = 1
    Sharpness of sigmoid soft masks M_S and M_B in Eqs. 10-11; set by hand.
  • lambda_deep / lambda_aux = 1 / 1
    Loss weights in Eq. 19; lambda_deep is analyzed in Fig. 5 and set to 1; lambda_aux is fixed at 1.
  • C' (projector channels) = not reported
    Output channel count of the Decision Evidence Projector (Eq. 4); architecture hyperparameter not specified in the paper.
  • Mamba configuration = not reported
    Dimensions, layers, and state sizes of the EG-Mamba module are not specified; essential for re-implementation.
axioms (5)
  • standard math Mamba selective state-space models provide a learnable long-range sequence operator.
    Eq. 9 uses Mamba M(·) as the state transition operator; the paper relies on [15] as a black box.
  • domain assumption Mesoscopic forensic traces (8-channel map from Mesorch) capture manipulation evidence.
    The entire pipeline conditions on F extracted from Mesorch (Sec. IV-C); if these traces miss AI-generated manipulations, the update process has no useful evidence to guide it.
  • domain assumption AI-generated forgeries leave subtle but detectable forensic traces.
    Core motivation in the Introduction; if this is false, progressive updating from weak evidence has nothing to exploit.
  • ad hoc to paper Shannon entropy of the initial one-shot map is a stable ambiguity prior.
    Eq. 5 computes U once from P0 and keeps it fixed (Sec. III-D1); the paper asserts rather than derives that this is a good prior for cautious updating.
  • ad hoc to paper Region-aware max/min preservation improves localization.
    Eqs. 13-15 encode a design choice—preserve confident regions, revise uncertain ones—that is validated empirically but not theoretically grounded.

pith-pipeline@v1.3.0-daily-deepseek · 17405 in / 12499 out tokens · 120103 ms · 2026-08-03T12:33:05.410437+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Progressive Decision-Making for Localizing Open-Ended AI-Generated Image Forgeries." pith.science (2026). https://pith.science/paper/WJY65IML

@misc{pith2026260729156,
  author       = {Pith},
  title        = {Pith review of: Progressive Decision-Making for Localizing Open-Ended AI-Generated Image Forgeries},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WJY65IML}},
  note         = {Machine review of arXiv:2607.29156}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

AI-generated image forgeries are becoming increasingly realistic and difficult to characterize with fixed manipulation patterns. As generative models continue to evolve, it is impractical to expect a localization model to exhaustively learn all possible forgery appearances from large-scale training data alone. Nevertheless, many AI-generated forgeries still leave subtle forensic traces, although these cues are often weak and unevenly reliable across regions. Therefore, robust localization requires not only extracting informative forensic traces, but also making reliable decisions from incomplete and ambiguous evidence. In this paper, we move beyond static one-shot prediction and reformulate final forgery localization as an adaptive sequential decision-updating process, where the localization map is treated as an intermediate state rather than a fixed output. Rather than producing the final mask via one-shot pixel-wise prediction, our method progressively updates the localization state guided by available evidence, uncertainty, and boundary conditions. Specifically, we first transform mesoscopic traces into compact decision evidence via a lightweight decision evidence projector, and then introduce Evidence-Guided Mamba (EG-Mamba) to perform uncertainty- and boundary-aware state updating. This design allows reliable manipulated and background regions to be preserved, while ambiguous regions are cautiously revised according to the available evidence. Extensive experiments on both conventional and AI-generated manipulation benchmarks validate the effectiveness of the proposed method. Notably, even when trained only on conventional manipulation data, our method brings larger gains on unseen AI-generated forgeries, indicating that progressive decision-updating is especially useful for heterogeneous and hard-to-exhaustively-learn manipulation traces.

Figures

Figures reproduced from arXiv: 2607.29156 by Jingyi Hou, Leyu Zhou, Xiaoxia Chen, Zhichuang Wang, Zhijie Liu.

Figure 1
Figure 1. Figure 1: Motivation of the proposed progressive decision [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed progressive forgery localization framework. Given an input image [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Analysis of the number of update steps T. F1-score and Permute F1-score are shown with the left and right y￾axes, respectively [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: Parameter analysis of the region-aware constrained [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison with representative image forgery localization methods. The first three rows are from AutoSplice, [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of the progressive localization states. The first three rows are from CASIAv1, and the last three rows [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗

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

46 extracted references · 1 linked inside Pith

  1. [1]

    Image manipulation detec- tion by multi-view multi-scale supervision,

    X. Chen, C. Dong, J. Ji, J. Cao, and X. Li, “Image manipulation detec- tion by multi-view multi-scale supervision,” inInternational Conference on Computer Vision. IEEE, 2021, pp. 14 165–14 173

    2021

  2. [2]

    Mvss-net: Multi- view multi-scale supervised networks for image manipulation detection,

    C. Dong, X. Chen, R. Hu, J. Cao, and X. Li, “Mvss-net: Multi- view multi-scale supervised networks for image manipulation detection,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 45, no. 3, pp. 3539–3553, 2023

    2023

  3. [3]

    Unionformer: Unified-learning transformer with multi-view representation for image manipulation detection and localization,

    S. Li, W. Ma, J. Guo, S. Xu, B. Li, and X. Zhang, “Unionformer: Unified-learning transformer with multi-view representation for image manipulation detection and localization,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2024, pp. 12 523– 12 533

    2024

  4. [4]

    Mantra-net: Manipulation tracing network for detection and localization of image forgeries with anomalous features,

    Y . Wu, W. AbdAlmageed, and P. Natarajan, “Mantra-net: Manipulation tracing network for detection and localization of image forgeries with anomalous features,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 9543–9552

    2019

  5. [5]

    Cat-net: Compression artifact tracing network for detection and localization of image splicing,

    M. Kwon, I. Yu, S. Nam, and H. Lee, “Cat-net: Compression artifact tracing network for detection and localization of image splicing,” in IEEE Winter Conference on Applications of Computer Vision, 2021, pp. 375–384

    2021

  6. [6]

    Learning JPEG compression artifacts for image manipulation detection and localization,

    M. Kwon, S. Nam, I. Yu, H. Lee, and C. Kim, “Learning JPEG compression artifacts for image manipulation detection and localization,” Int. J. Comput. Vis., vol. 130, no. 8, pp. 1875–1895, 2022

    2022

  7. [7]

    Edge-aware regional message passing controller for image forgery localization,

    D. Li, J. Zhu, M. Wang, J. Liu, X. Fu, and Z. Zha, “Edge-aware regional message passing controller for image forgery localization,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 8222–8232. SUBMISSION MANUSCRIPT 13

    2023

  8. [8]

    EAN: edge-aware network for image manipulation localization,

    Y . Chen, H. Cheng, H. Wang, X. Liu, F. Chen, F. Li, X. Zhang, and M. Wang, “EAN: edge-aware network for image manipulation localization,”IEEE Trans. Circuits Syst. Video Technol., vol. 35, no. 2, pp. 1591–1601, 2025

    2025

  9. [9]

    Objectformer for image manipulation detection and localization,

    J. Wang, Z. Wu, J. Chen, X. Han, A. Shrivastava, S. Lim, and Y . Jiang, “Objectformer for image manipulation detection and localization,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 2354–2363

    2022

  10. [10]

    Hdf-net: Capturing homogeny difference features to localize the tampered image,

    R. Han, X. Wang, N. Bai, Y . Wang, J. Hou, and J. Xue, “Hdf-net: Capturing homogeny difference features to localize the tampered image,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 46, no. 12, pp. 10 005– 10 020, 2024

    2024

  11. [11]

    Learnable frequency decomposition for image forgery detection and localization,

    D. Li, J. Zhu, Y . Liu, X. Lu, X. Fu, J. Liu, A. Liu, and Z. Zha, “Learnable frequency decomposition for image forgery detection and localization,” inInternational Joint Conference on Artificial Intelligence, 2025, pp. 1359–1367

    2025

  12. [12]

    Mesoscopic insights: orchestrating multi-scale & hybrid architecture for image manipulation localization,

    X. Zhu, X. Ma, L. Su, Z. Jiang, B. Du, X. Wang, Z. Lei, W. Feng, C.- M. Pun, and J.-Z. Zhou, “Mesoscopic insights: orchestrating multi-scale & hybrid architecture for image manipulation localization,” inAAAI Conference on Artificial Intelligence, 2025, pp. 11 022–11 030

    2025

  13. [13]

    Uncertainty-guided learning for improving image manipulation detection,

    K. Ji, F. Chen, X. Guo, Y . Xu, J. Wang, and J. Chen, “Uncertainty-guided learning for improving image manipulation detection,” inInternational Conference on Computer Vision, 2023, pp. 22 456–22 465

    2023

  14. [14]

    Employing reinforcement learning to construct a decision-making environment for image forgery localization,

    R. Peng, S. Tan, X. Mo, B. Li, and J. Huang, “Employing reinforcement learning to construct a decision-making environment for image forgery localization,”IEEE Trans. Inf. Forensics Secur., vol. 19, pp. 4820–4834, 2024

    2024

  15. [15]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” inConference on Language Modeling, 2024

    2024

  16. [16]

    Collaborative transformers with multi-level forensic attention for image manipulation localization,

    J. Zhang, W. Feng, S. Wang, F. Kou, H. Yu, and S. Niu, “Collaborative transformers with multi-level forensic attention for image manipulation localization,” inAAAI Conference on Artificial Intelligence, S. Koenig, C. Jenkins, and M. E. Taylor, Eds., 2026, pp. 12 556–12 563

    2026

  17. [17]

    Pixel- inconsistency modeling for image manipulation localization,

    C. Kong, A. Luo, S. Wang, H. Li, A. Rocha, and A. C. Kot, “Pixel- inconsistency modeling for image manipulation localization,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 47, no. 6, pp. 4455–4472, 2025

    2025

  18. [18]

    Dirloc: Disentanglement representation learning for robust image forgery localization,

    Z. Sheng, Z. Qu, W. Lu, X. Cao, and J. Huang, “Dirloc: Disentanglement representation learning for robust image forgery localization,”IEEE Trans. Dependable Secur. Comput., vol. 22, no. 3, pp. 2841–2854, 2025

    2025

  19. [19]

    Generate, segment, and refine: Towards generic manipulation segmentation,

    P. Zhou, B. Chen, X. Han, M. Najibi, A. Shrivastava, S. Lim, and L. Davis, “Generate, segment, and refine: Towards generic manipulation segmentation,” inAAAI Conference on Artificial Intelligence, 2020, pp. 13 058–13 065

    2020

  20. [20]

    M2sformer: Multi-spectral and multi-scale attention with edge-aware difficulty guidance for image forgery localization,

    J.-H. Nam, D.-H. Moon, and S.-C. Lee, “M2sformer: Multi-spectral and multi-scale attention with edge-aware difficulty guidance for image forgery localization,” inInternational Conference on Computer Vision, 2025, pp. 15 927–15 938

    2025

  21. [21]

    Recurrent models of visual attention,

    V . Mnih, N. Heess, A. Graves, and K. Kavukcuoglu, “Recurrent models of visual attention,” inAdvances in Neural Information Processing Systems, 2014, pp. 2204–2212

    2014

  22. [22]

    Glance and focus networks for dynamic visual recognition,

    G. Huang, Y . Wang, K. Lv, H. Jiang, W. Huang, P. Qi, and S. Song, “Glance and focus networks for dynamic visual recognition,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 45, no. 4, pp. 4605–4621, 2023

    2023

  23. [23]

    Polyformer: Referring image segmentation as sequential polygon generation,

    J. Liu, H. Ding, Z. Cai, Y . Zhang, R. K. Satzoda, V . Mahadevan, and R. Manmatha, “Polyformer: Referring image segmentation as sequential polygon generation,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 18 653–18 663

    2023

  24. [24]

    Ugee- net: Uncertainty-guided and edge-enhanced network for image splicing localization,

    Q. Hao, R. Ren, S. Niu, K. Wang, M. Wang, and J. Zhang, “Ugee- net: Uncertainty-guided and edge-enhanced network for image splicing localization,”Neural Networks, vol. 178, p. 106430, 2024

    2024

  25. [25]

    Inpdiffusion: Image inpainting localization via conditional diffusion models,

    K. Wang, S. Niu, Q. Hao, and J. Zhang, “Inpdiffusion: Image inpainting localization via conditional diffusion models,” inAAAI Conference on Artificial Intelligence, T. Walsh, J. Shah, and Z. Kolter, Eds., 2025, pp. 7771–7779

    2025

  26. [26]

    Forgdiffuser: General image forgery localization with diffusion models,

    M. Wang, S. Niu, and J. Zhang, “Forgdiffuser: General image forgery localization with diffusion models,” inInternational Joint Conference on Artificial Intelligence, 2025, pp. 1954–1962

    2025

  27. [27]

    Image forgery localization with state space models,

    Z. Lou, G. Cao, K. Guo, S. Weng, and L. Yu, “Image forgery localization with state space models,”IEEE Signal Process. Lett., vol. 32, pp. 1590– 1594, 2025

    2025

  28. [28]

    A lightweight and effective image tampering localization network with vision mamba,

    K. Guo, G. Cao, Z. Lou, X. Huang, and J. Liu, “A lightweight and effective image tampering localization network with vision mamba,” IEEE Signal Process. Lett., vol. 32, pp. 2179–2183, 2025

    2025

  29. [29]

    Towards modern image manipulation localization: A large-scale dataset and novel methods,

    C. Qu, Y . Zhong, C. Liu, G. Xu, D. Peng, F. Guo, and L. Jin, “Towards modern image manipulation localization: A large-scale dataset and novel methods,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 10 781–10 790

    2024

  30. [30]

    GIM: A million-scale benchmark for generative image manipulation detection and localization,

    Y . Chen, X. Huang, Q. Zhang, W. Li, M. Zhu, Q. Yan, S. Li, H. Chen, H. Hu, J. Yang, W. Liu, and J. Hu, “GIM: A million-scale benchmark for generative image manipulation detection and localization,” inAAAI Conference on Artificial Intelligence, T. Walsh, J. Shah, and Z. Kolter, Eds., 2025, pp. 2311–2319

    2025

  31. [31]

    Detective sam: Adaptive ai-image forgery localization,

    G. Lek, N. Van Schaik, C. Zhu, P.-Y . Chen, R. Birke, and L. Y . Chen, “Detective sam: Adaptive ai-image forgery localization,” inInternational Conference on Learning Representations, 2026

    2026

  32. [32]

    Fakeshield: Explainable image forgery detection and localization via multi-modal large language models,

    Z. Xu, X. Zhang, R. Li, Z. Tang, Q. Huang, and J. Zhang, “Fakeshield: Explainable image forgery detection and localization via multi-modal large language models,” inInternational Conference on Learning Rep- resentations, 2025

    2025

  33. [33]

    Manipshield: A unified framework for image manipulation detection, localization and explanation,

    Z. Xu, H. Duan, X. Wang, Z. Cai, K. Zhang, Q. Hu, J. Liu, X. Min, and G. Zhai, “Manipshield: A unified framework for image manipulation detection, localization and explanation,”CoRR, vol. abs/2511.14259,

    arXiv

  34. [34]

    Imdl-benco: A comprehensive benchmark and codebase for image manipulation detection & localization,

    X. Ma, X. Zhu, L. Su, B. Du, Z. Jiang, B. Tong, Z. Lei, X. Yang, C.-M. Pun, J. Lvet al., “Imdl-benco: A comprehensive benchmark and codebase for image manipulation detection & localization,”Advances in Neural Information Processing Systems, vol. 37, pp. 134 591–134 613, 2024

    2024

  35. [35]

    Casia image tampering detection eval- uation database,

    J. Dong, W. Wang, and T. Tan, “Casia image tampering detection eval- uation database,” inIEEE China Summit and International Conference on Signal and Information Processing, 2013, pp. 422–426

    2013

  36. [36]

    The point where reality meets fantasy: Mixed adversarial generators for image splice detection,

    V . V . Kniaz, V . A. Knyaz, and F. Remondino, “The point where reality meets fantasy: Mixed adversarial generators for image splice detection,” inAdvances in Neural Information Processing Systems, H. M. Wallach, H. Larochelle, A. Beygelzimer, F. d’Alch ´e-Buc, E. B. Fox, and R. Garnett, Eds., 2019, pp. 215–226

    2019

  37. [37]

    IMD2020: A large-scale annotated dataset tailored for detecting manipulated images,

    A. Novoz ´amsk´y, B. Mahdian, and S. Saic, “IMD2020: A large-scale annotated dataset tailored for detecting manipulated images,” inIEEE Winter Applications of Computer Vision Workshops, 2020, pp. 71–80

    2020

  38. [38]

    MFC datasets: Large-scale benchmark datasets for media forensic challenge evalua- tion,

    H. Guan, M. Kozak, E. Robertson, Y . Lee, A. N. Yates, A. Delgado, D. Zhou, T. Kheyrkhah, J. Smith, and J. G. Fiscus, “MFC datasets: Large-scale benchmark datasets for media forensic challenge evalua- tion,” inIEEE Winter Applications of Computer Vision Workshops, 2019, pp. 63–72

    2019

  39. [39]

    Autosplice: A text- prompt manipulated image dataset for media forensics,

    S. Jia, M. Huang, Z. Zhou, Y . Ju, J. Cai, and S. Lyu, “Autosplice: A text- prompt manipulated image dataset for media forensics,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 893– 903

    2023

  40. [40]

    Sagi: Semantically aligned and uncertainty guided ai image inpainting,

    P. Giakoumoglou, D. Karageorgiou, S. Papadopoulos, and P. C. Petran- tonakis, “Sagi: Semantically aligned and uncertainty guided ai image inpainting,” inInternational Conference on Computer Vision, 2025, pp. 16 090–16 101

    2025

  41. [41]

    A convnet for the 2020s,

    Z. Liu, H. Mao, C. Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A convnet for the 2020s,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 11 966–11 976

    2022

  42. [42]

    Segformer: Simple and efficient design for semantic segmentation with transformers,

    E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. ´Alvarez, and P. Luo, “Segformer: Simple and efficient design for semantic segmentation with transformers,” inAdvances in Neural Information Processing Systems, 2021, pp. 12 077–12 090

    2021

  43. [43]

    Pscc-net: Progressive spatio- channel correlation network for image manipulation detection and localization,

    X. Liu, Y . Liu, J. Chen, and X. Liu, “Pscc-net: Progressive spatio- channel correlation network for image manipulation detection and localization,”IEEE Trans. Circuits Syst. Video Technol., vol. 32, no. 11, pp. 7505–7517, 2022

    2022

  44. [44]

    Trufor: Leveraging all-round clues for trustworthy image forgery detection and localization,

    F. Guillaro, D. Cozzolino, A. Sud, N. Dufour, and L. Verdoliva, “Trufor: Leveraging all-round clues for trustworthy image forgery detection and localization,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 20 606–20 615

    2023

  45. [45]

    Iml-vit: Benchmarking image manipulation localization by vision transformer,

    X. Ma, B. Du, Z. Jiang, A. Y . A. Hammadi, and J. Zhou, “Iml-vit: Benchmarking image manipulation localization by vision transformer,” arXiv preprint arXiv:2307.14863, 2023

    Pith/arXiv arXiv 2023

  46. [2025]

    Available: https://doi.org/10.48550/arXiv.2511.14259

    [Online]. Available: https://doi.org/10.48550/arXiv.2511.14259

    doi:10.48550/arxiv.2511.14259