Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

IMDPrompter: Adapting SAM to Image Manipulation Detection by Cross-View Automated Prompt Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read The paper shows that the Segment Anything Model can be converted into a fully automated image manipulation detector by learning prompts from four complementary views, and that this system, trained only on CASIAv2, outperforms prior…

desk verdict A solid, well-ablated SAM-based forensics paper whose headline Optimal Prompt Selection mechanism is trained with ground-truth oracle selection but at inference quietly falls back to a fixed ensemble mean — the mismatch is never analyzed, and a backbone contradiction (VIT-L vs SAM-H) needs resolving. read the letter →

arxiv 2502.02454 v4 pith:DNFYC4EC submitted 2025-02-04 cs.CV

Quan Zhang , Yuxin Qi , Xi Tang , Jinwei Fang , Xi Lin , Ke Zhang , Chun Yuan This is my paper · ORCID
classification cs.CV
keywords imagemanipulationdetectionforgerylocalizationSegmentAnythingModelpromptlearningcross-viewfeaturefusionNoiseprintoptimalselectioncross-datasetgeneralization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper sets out to prove that a general-purpose segmentation foundation model, SAM, can be repurposed into a fully automatic image-manipulation detector, removing the manual point-and-box prompts SAM normally requires. The proposed IMDPrompter learns prompts from four views of the same image: plain RGB plus three noise-focused representations (SRM, Bayer-filter residuals, and Noiseprint). By selecting the best-view prompt at training time, pulling all views toward it, and feeding the fused features and prompt into SAM's mask decoder, the model localizes spliced, copy-moved, and inpainted regions. Trained only on CASIAv2, it reports the best pixel-level and image-level F1 on CASIA, COVER, Columbia, IMD2020, and NIST16, with an average pixel F1 of 81.36% at the best threshold and 59.78% at a fixed 0.5 threshold. If the results hold, a frozen, prompt-driven foundation model can replace specialist forensic networks and generalize across datasets without manual threshold tuning.

What carries the argument

The load-bearing mechanism is the cross-view automated prompt: instead of asking a human for points or boxes, four lightweight segmenters, one per view, produce mask probability maps, and the best one is converted into a mask-plus-box prompt for SAM's prompt encoder. Two supporting mechanisms keep that prompt reliable: the Cross-View Prompt Consistency loss trains each view to converge toward the selected optimal mask, and the Cross-View Feature Perception module fuses all four views' features into SAM's decoder. The Prompt Mixing Module merges the selected prompt embedding with the fused features before the mask decoder produces the final localization. The net effect is that SAM's frozen, billion-mask pretraining is carried into forensics entirely through learned prompts and a small trainable prompt-and-feature pathway.

What would settle it

Construct a test set of manipulated images in which the Noiseprint view is the only accurate view while the other three views are confidently wrong, for example images with unusual sensor noise or heavy denoising that breaks the fingerprint. If IMDPrompter's pixel F1 collapses to the level of the poor views under its fixed ensemble average, the central assumption that ensemble averaging preserves the optimal view is refuted.

Watch

Extended reading notes

Core claim

The central discovery is that SAM's mask decoder can be driven by learned prompts derived from semantic-agnostic noise views, and that this works better than training a specialist segmentation network. IMDPrompter generates candidate masks from four views, and during training the Optimal Prompt Selection module picks the candidate with the lowest segmentation loss against the ground-truth mask, converts it into a mask-plus-box prompt for SAM, and a Cross-View Prompt Consistency loss forces all views to agree with that optimal choice. At inference, when labels are unavailable, the selection module falls back to the simple average of the four view predictions. The paper reports that noise views, especially Noiseprint, are chosen as optimal far more often than RGB (46.3% versus 13.9% of training cases), and that the full system beats the strongest prior method, MVSS-Net, by large margins on both pixel-level and image-level metrics. This is presented as the first application of SAM to image manipulation detection and localization.

Load-bearing premise

The method assumes that at inference, where no labels exist and the per-image optimal view cannot be selected, the fixed average of the four views' masks remains a reliable proxy for the trained optimal-prompt teacher on out-of-distribution data.

Editorial extensions

If this is right

  • If the reported numbers hold, improving the detector becomes a matter of improving prompt generation rather than retraining a specialist forensic backbone.
  • The fixed-threshold pixel F1 of 59.78% versus 41.90% for MVSS-Net implies the method is less sensitive to the deployment threshold, which matters in practice where the operating point is not known in advance.
  • Because Noiseprint is selected as the optimal view in 46.3% of training cases, the noise-fingerprint view carries much of the cross-dataset generalization, supporting the paper's claim that semantic-agnostic views are essential for out-of-distribution performance.
  • The same prompt-learning recipe is claimed to transfer to other visual foundation models beyond SAM, suggesting a general template for adapting segmentation foundation models to forensic tasks.
  • The robustness tables show the method degrades more gracefully than baselines under JPEG compression and Gaussian blur, the kinds of operations images undergo during online dissemination.

Reading between the lines

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

  • The train-versus-inference mismatch of the optimal prompt selection module (per-image optimal during training, ensemble average at test time) is the most likely failure mode; a test-time adaptation that re-weights views by their internal consistency could close that gap, but the paper does not test this.
  • The dominance of the Noiseprint view suggests that adding further camera-model or sensor-noise fingerprints, or learning new noise views end-to-end, may improve generalization more than adding additional RGB-based supervision.
  • Because the method requires full pixel-level labels, a weakly supervised variant is the natural next step; the paper names this direction in its limitations but does not explore it.
  • The reported 45.3% composite-F1 improvement over the next best method is partly inflated by the weak image-level performance of prior baselines; on generative-image edits (the CocoGlide benchmark) the gains are smaller, so the practical advantage on AI-generated forgeries is narrower than the headline number.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes IMDPrompter, a framework that adapts the Segment Anything Model (SAM) to image manipulation detection (IMD) and localization by learning automated cross-view prompts. Four views (RGB, SRM, Bayer, Noiseprint) are processed by separate segmenters; an Optimal Prompt Selection (OPS) module picks the best view prediction during training using ground-truth labels, a Cross-View Prompt Consistency (CPC) loss encourages views to align with the selected optimum, a Cross-View Feature Perception (CFP) module fuses features, and a Prompt Mixing Module (PMM) integrates prompt embeddings for SAM's mask decoder. The model is trained only on CASIAv2 and evaluated on CASIAv1, COVER, Columbia, IMD2020, and NIST16, reporting state-of-the-art image-level and pixel-level F1 scores, together with ablations and robustness analyses.

Significance. If the results are reliable, the paper makes a valuable contribution by demonstrating that SAM's strong segmentation priors can be harnessed for IMD through automated prompt learning, with particular promise for out-of-distribution generalization. The manuscript is commendably concrete: the loss functions are clearly specified, the ablation study isolates each component, the evaluation covers five benchmarks, and the appendix includes parameter settings, view-selection statistics, robustness tables, and complexity comparisons. However, the significance is conditional on two load-bearing issues: the OPS module's oracle-based training objective is replaced at inference by a fixed ensemble average without any analysis of the mismatch, and the headline pixel-level results are reported under a best-threshold protocol that selects thresholds using test labels. These issues weaken the direct support for the central SOTA claim.

major comments (3)
  1. [Section 3.2, Eq. (7); Section 3.6; Tables 3-4] The OPS module is trained with an oracle selector: Eq. (7) chooses Popt = argmin over the four view predictions and their ensemble using ground-truth labels G, and the CPC loss in Eq. (10) pushes each view toward that oracle choice. At inference, Section 3.6 states that the OPS module 'defaults to selecting PEns', i.e., a fixed unweighted average of the four view maps. The paper does not quantify how often PEns coincides with the training-time oracle, how close PEns is to the oracle on out-of-distribution data, or how PEns compares with the best single view at inference. The OPS ablation gains in Tables 3 and 4 (e.g., +4.5/+5.8/+5.2 on CASIA I-F1/P-F1/C-F1 and +44.8/+21.6/+36.2 on COVER) therefore measure a training-time oracle behavior that is not operationalized at test time; the reported SOTA results are for an ensemble-average model, not for the proposed optimal-prompt selector. This is a central claim that requires either an inference-time selection mechanism or a quantitative demonstration that PEns is a reliable proxy for the oracle optimum.
  2. [Table 1 and Appendix Table 10] The headline pixel-level F1 of 81.36% (average across five datasets) is reported under the 'Best threshold' protocol, where the threshold is selected on test labels, whereas the fixed-threshold 0.5 results are substantially lower (average 59.78%). Because the paper's conclusions and the 45.3% improvement claim in Appendix Table 10 rely on the harmonic mean that embeds these best-threshold P-F1 values, the practical advantage of IMDPrompter is overstated relative to what is achievable without test-label access. The authors should either present fixed-threshold P-F1 as the primary comparison or explicitly justify why best-threshold selection is the appropriate protocol and discuss the optimistic bias it introduces.
  3. [Section 4 (Implementation Details) and Appendix Table 15] The main text states 'we consistently use the VIT-L backbone of SAM,' but Appendix Table 15 says 'the IMDPrompter mentioned in the main text is based on SAM-H.' This is a direct factual contradiction that prevents a reader from knowing which backbone produced Tables 1 and 2. The inconsistency must be resolved for the reported results to be reproducible and for the complexity comparison in Table 15 to be interpretable.
minor comments (5)
  1. [Abstract] The sentence 'we develops a cross-view prompt learning paradigm' should be 'we develop a cross-view prompt learning paradigm.'
  2. [Section 4.1] The citation 'As shown in Figure 10' should refer to the main-text qualitative figure (Figure 4), since Figures 8-10 are in the appendix.
  3. [Appendix Section 7.4] The sentence 'As shown in Table 3, IMDPrompter achieves the best performance in all settings' should reference Table 10, not Table 3.
  4. [Appendix Section 7.4] There is a typo in the SAM-based methods comparison: 'CVOER' should be 'COVER.' Also, 'iimage' appears in the abstract and in the appendix; both should be 'image.'
  5. [Table 15] The table lists both IMDPrompter and IMDPrompter* but the main text does not explain that IMDPrompter* uses Mobile SAM until the appendix; this should be clarified earlier to avoid confusion about which model produces the main SOTA numbers.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: training uses ground-truth supervision only; at inference the method defaults to the ensemble prompt, so test predictions are not constructed from test labels.

full rationale

IMDPrompter is an empirical pipeline that adapts a frozen SAM with learned prompt generators; there is no formal derivation that could reduce to its own inputs. During training, Eq. 7 selects the optimal mask probability Popt = argmin_{P in P} LSeg(P, G) using the ground-truth mask G, and Eq. 10 then trains each view toward that oracle-selected map. At inference, Section 3.6 explicitly states: "During the inference process of IMDPrompter, since there are no true labels, our OPS module defaults to selecting PEns for generating masks and bounding box prompts." Thus the reported test results are produced by the fixed ensemble prompt, not by a label-dependent selector. This is a train/inference mismatch worth investigating, because the paper does not quantify how often PEns coincides with the training-time oracle Popt on out-of-distribution data, but it is not circular: the final predictions are not constructed from the test labels. The best-threshold P-F1 in Table 1 is selected on test labels, but the paper also reports fixed-threshold (0.5) P-F1, and the composite C-F1 in Table 10 is computed from fixed-threshold values, so the central SOTA comparison is not an identity. The model is trained only on CASIAv2 and evaluated on held-out external benchmarks, and no load-bearing self-citation or imported uniqueness theorem is invoked. Therefore no circular step is present.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim relies on a few domain assumptions inherited from prior IMD literature: that noise views capture manipulation traces, that SAM features can be repurposed via prompts, and that the ensemble prediction is a suitable inference-time proxy for the ground-truth-optimal prompt. All are plausible and partially validated by ablations, but none is formally proven.

free parameters (2)
  • Loss weights lambda1, lambda2, lambda3 = lambda1=1.0, lambda2=0.1, lambda3=1.0
    Selected via hyperparameter analysis (Figure 7) to balance segmentation, consistency, and image-level losses; reported final values.
  • Best threshold per dataset for pixel F1 = Dataset-specific thresholds chosen to maximize F1 on test labels (CASIA, COVER, Columbia, IMD, NIST16)
    Table 1's 'best threshold' column selects the operating point on the test set, which is an oracle fit; the paper also reports a fixed threshold of 0.5.
assumptions (4)
  • domain assumption SAM's frozen image encoder features can be effectively conditioned by prompts derived from forensic noise views (SRM, Bayar, Noiseprint) to localize manipulations.
    The whole CFP/PMM design assumes object-agnostic SAM features transfer to tamper localization; Sections 3.4 and 3.5 build on this without direct verification beyond the final metrics.
  • ad hoc to paper At inference, the mean of the four view predictions PEns is a reasonable stand-in for the ground-truth-optimal prompt when labels are absent.
    Section 3.6 states OPS 'defaults to selecting PEns' during inference; no evidence is given that PEns approximates the per-sample optimal view on OOD data.
  • domain assumption SRM, Bayar, and Noiseprint filters provide complementary, semantic-agnostic traces of manipulation that are useful across datasets.
    Taken from prior IMD literature (ManTra-Net, MVSS-Net, TruFor); the paper relies on this for cross-dataset gains.
  • domain assumption Otsu thresholding and aggregation over pixels above the threshold is a valid image-level detector.
    Adopted from Zhai et al. (2023); used in Equations 17-18 without adaptation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of IMDPrompter: Adapting SAM to Image Manipulation Detection by Cross-View Automated Prompt Learning." pith.science (2026). https://pith.science/paper/DNFYC4EC

@misc{pith2026250202454,
  author       = {Pith},
  title        = {Pith review of: IMDPrompter: Adapting SAM to Image Manipulation Detection by Cross-View Automated Prompt Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DNFYC4EC}},
  note         = {Machine review of arXiv:2502.02454}
}
read the original abstract

Using extensive training data from SA-1B, the Segment Anything Model (SAM) has demonstrated exceptional generalization and zero-shot capabilities, attracting widespread attention in areas such as medical image segmentation and remote sensing image segmentation. However, its performance in the field of image manipulation detection remains largely unexplored and unconfirmed. There are two main challenges in applying SAM to image manipulation detection: a) reliance on manual prompts, and b) the difficulty of single-view information in supporting cross-dataset generalization. To address these challenges, we develops a cross-view prompt learning paradigm called IMDPrompter based on SAM. Benefiting from the design of automated prompts, IMDPrompter no longer relies on manual guidance, enabling automated detection and localization. Additionally, we propose components such as Cross-view Feature Perception, Optimal Prompt Selection, and Cross-View Prompt Consistency, which facilitate cross-view perceptual learning and guide SAM to generate accurate masks. Extensive experimental results from five datasets (CASIA, Columbia, Coverage, IMD2020, and NIST16) validate the effectiveness of our proposed method.

Figures

Figures reproduced from arXiv: 2502.02454 by the authors.

Figure 1
Figure 1. Improvements of proposed IMDPrompter over existing pipelines of baselines: (a). Automated prompt learning without the need for manual input. (b). Flexible integration of semantic-agnostic information crucial for Image Manipulation Detection. To activate SAM’s IMD capabilities, we propose a cross-view automated prompt learning paradigm called IMDPrompter, aimed at learning how to generate prompts to enhance SAM’s fun… view at source ↗
Figure 2
Figure 2. Overall framework of IMDPrompter. The prompter part consists of four views: RGB, SRM, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Architecture of the Cross-view Feature Perception (CFP) unit. To enhance the segmentation accuracy of SAM, we propose a Cross-View Feature Perception (CFP) module that integrates features from RGB, SRM, Bayer, and Noiseprint views. This module enriches the feature representations for SAM’s de￾coding process, as shown in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Visualization Comparison of Image Manipulation Detection and Localization Results [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Robustness Analysis Against JPEG Compression and Gaussian Blur Interference [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Visualization of Metric Data Comparison between IMDPrompter and State-of-the-Art [PITH_FULL_IMAGE:figures/full_fig_p020_6.png]
Figure 7
Figure 7. Figure 7: Hyperparametric analysis of λ1, λ2, and λ3 Proportion of Four Views Selected as Optimal Prompts. As shown in [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: Some qualitative results, compared with the state-of-the-art [PITH_FULL_IMAGE:figures/full_fig_p023_8.png]
Figure 9
Figure 9. Figure 9: Some qualitative results, compared with the state-of-the-art [PITH_FULL_IMAGE:figures/full_fig_p024_9.png]
Figure 10
Figure 10. Figure 10: Some qualitative results, compared with the state-of-the-art [PITH_FULL_IMAGE:figures/full_fig_p025_10.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. $K^2$VAE: A Koopman-Kalman Enhanced Variational AutoEncoder for Probabilistic Time Series Forecasting

    cs.LG 2025-05 conditional novelty 6.0 of 10

    Combining a learned Koopman linearization with a learned Kalman filter inside a VAE produces a probabilistic forecaster that beats existing methods on most tested short- and long-horizon datasets.

Reference graph

Works this paper leans on

38 extracted references · 14 canonical work pages · cited by 1 Pith paper

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774,

    arXiv

  2. [3]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901,

    1901

  3. [5]

    Splicebuster: A new blind image splicing detector

    Davide Cozzolino, Giovanni Poggi, and Luisa Verdoliva. Splicebuster: A new blind image splicing detector. In 2015 IEEE International Workshop on Information Forensics and Security (WIFS), pp. 1–6. IEEE,

    work page 2015

  4. [9]

    Table 10 and Figure 6 shows the overall performance of pixel-level and image-level manipulation detection

    Overall Performance in Detection and Localization. Table 10 and Figure 6 shows the overall performance of pixel-level and image-level manipulation detection. We use the harmonic mean of image-level detection F1 and pixel-level localization F1, denoted as C-F1, as our overall performance metric. As shown in Table 3, IMDPrompter achieves the best performanc...

    work page 2019

  5. [14]

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

    12 Published as a conference paper at ICLR 2025 Myung-Joon Kwon, In-Jae Yu, Seung-Hun Nam, and Heung-Kyu Lee. Cat-net: Compression artifact tracing network for detection and localization of image splicing. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp. 375–384,

    work page 2025

  6. [15]

    The power of scale for parameter-efficient prompt tuning

    Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691,

    arXiv

  7. [16]

    Foundts: Comprehensive and unified benchmarking of foundation models for time series forecasting

    Zhe Li, Xiangfei Qiu, Peng Chen, Yihang Wang, Hanyin Cheng, Yang Shu, Jilin Hu, Chenjuan Guo, Aoying Zhou, Qingsong Wen, et al. Foundts: Comprehensive and unified benchmarking of foundation models for time series forecasting. arXiv preprint arXiv:2410.11802,

    arXiv

  8. [17]

    Sgdr: Stochastic gradient descent with warm restarts

    Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983,

    arXiv

Show all 38 references
  1. [20]

    Codev-bench: How do llms understand developer-centric code completion? arXiv preprint arXiv:2410.01353, 2024a

    Zhenyu Pan, Rongyu Cao, Yongchang Cao, Yingwei Ma, Binhua Li, Fei Huang, Han Liu, and Yongbin Li. Codev-bench: How do llms understand developer-centric code completion? arXiv preprint arXiv:2410.01353, 2024a. Zhenyu Pan, Haozheng Luo, Manling Li, and Han Liu. Conv-coa: Improvi...

    arXiv

  2. [22]

    Sam 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨adle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714,

    arXiv

  3. [23]

    Autosam: Adapting sam to medical images by overloading the prompt encoder

    Tal Shaharabany, Aviad Dahan, Raja Giryes, and Lior Wolf. Autosam: Adapting sam to medical images by overloading the prompt encoder. arXiv preprint arXiv:2306.06370,

    arXiv

  4. [24]

    Cnn features off-the-shelf: an astounding baseline for recognition

    14 Published as a conference paper at ICLR 2025 Ali Sharif Razavian, Hossein Azizpour, Josephine Sullivan, and Stefan Carlsson. Cnn features off-the-shelf: an astounding baseline for recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition ...

    2025

  5. [26]

    Self-consistency improves chain of thought reasoning in language models

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdh- ery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171, 2022a. Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, R...

    arXiv

  6. [27]

    Task groupings regularization: Data-free meta-learning with heterogeneous pre-trained models

    Yongxian Wei, Zixuan Hu, Li Shen, Zhenyi Wang, Yu Li, Chun Yuan, and Dacheng Tao. Task groupings regularization: Data-free meta-learning with heterogeneous pre-trained models. arXiv preprint arXiv:2405.16560, 2024a. Yongxian Wei, Zixuan Hu, Zhenyi Wang, Li Shen, Chun Yuan, and...

    arXiv

  7. [29]

    Medical sam adapter: Adapting segment anything model for medical image segmentation

    Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yueming Jin. Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620,

    arXiv

  8. [30]

    Constrained r-cnn: A general image manipulation detection model

    Chao Yang, Huizhou Li, Fangting Lin, Bin Jiang, and Hao Zhao. Constrained r-cnn: A general image manipulation detection model. In 2020 IEEE International conference on multimedia and expo (ICME), pp. 1–6. IEEE,

    2020

  9. [31]

    Multi-modality image manipula- tion detection

    15 Published as a conference paper at ICLR 2025 Chao Yang, Zhiyu Wang, Huawei Shen, Huizhou Li, and Bin Jiang. Multi-modality image manipula- tion detection. In 2021 IEEE International conference on multimedia and expo (ICME), pp. 1–6. IEEE,

    2025

  10. [32]

    Customized segment anything model for medical image segmentation

    Kaidong Zhang and Dong Liu. Customized segment anything model for medical image segmentation. arXiv preprint arXiv:2304.13785,

    arXiv

  11. [33]

    Can mllms guide weakly-supervised temporal action localization tasks? arXiv preprint arXiv:2411.08466,

    Quan Zhang and Yuxin Qi. Can mllms guide weakly-supervised temporal action localization tasks? arXiv preprint arXiv:2411.08466,

    arXiv

  12. [34]

    IMDPrompter: Adapting SAM to image manipulation detection by cross-view automated prompt learning

    Quan Zhang, Yuxin Qi, Xi Tang, Jinwei Fang, Xi Lin, Ke Zhang, and Chun Yuan. IMDPrompter: Adapting SAM to image manipulation detection by cross-view automated prompt learning. In The Thirteenth International Conference on Learning Representations , 2025a. URL https: //openrevi...

    arXiv

  13. [35]

    Generate, segment and replace: Towards generic manipulation segmentation

    Peng Zhou, B Chen, Xintong Han, Mahyar Najibi, and Larry S Davis. Generate, segment and replace: Towards generic manipulation segmentation. arXiv preprint arXiv:1811.09729, 2018a. Peng Zhou, Xintong Han, Vlad I Morariu, and Larry S Davis. Learning rich features for image manip...

    arXiv

  14. [36]

    We have some limitations in our method

    16 Published as a conference paper at ICLR 2025 7 A PPENDIX 7.1 L IMITATIONS AND BROADER IMPACTS Limitations. We have some limitations in our method. First, it cannot detect completely generated images. Second, training IMDPrompter requires complete pixel-level supervision. In...

    2024

  15. [37]

    Based on these criteria, we compiled a list of nine published baselines, as follows: • H-LSTM Bappy et al

    adherence to a common evaluation protocol, where CASIAv2 is used for training and other public datasets for testing. Based on these criteria, we compiled a list of nine published baselines, as follows: • H-LSTM Bappy et al. (2019): Pretrained on a custom dataset of 65k process...

    2019

  16. [1975]

    Training language models to follow instructions with human feedback

    13 Published as a conference paper at ICLR 2025 Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances i...

    2025

  17. [2006]

    Multirc: Joint learning for time series anomaly prediction and detection with multi-scale reconstructive contrast

    Shiyan Hu, Kai Zhao, Xiangfei Qiu, Yang Shu, Jilin Hu, Bin Yang, and Chenjuan Guo. Multirc: Joint learning for time series anomaly prediction and detection with multi-scale reconstructive contrast. arXiv preprint arXiv:2410.15997,

    arXiv

  18. [2013]

    Dfvsr: Directional frequency video super-resolution via asymmetric and enhancement alignment network

    Shuting Dong, Feng Lu, Zhe Wu, and Chun Yuan. Dfvsr: Directional frequency video super-resolution via asymmetric and enhancement alignment network. In IJCAI, pp. 681–689, 2023a. Shuting Dong, Feng Lu, and Chun Yuan. Frequency reciprocal action and fusion for single image super...

    2023 arXiv

  19. [2014]

    Very deep convolutional networks for large-scale image recognition

    Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556,

    arXiv

  20. [2015]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805,

    arXiv

  21. [2016]

    Mobilenets: Efficient convolutional neural networks for mobile vision applica- tions

    Andrew G Howard. Mobilenets: Efficient convolutional neural networks for mobile vision applica- tions. arXiv preprint arXiv:1704.04861,

    arXiv

  22. [2017]

    Detecting image splicing using geometry invariants and camera characteristics consistency

    Yu-Feng Hsu and Shih-Fu Chang. Detecting image splicing using geometry invariants and camera characteristics consistency. In 2006 IEEE International Conference on Multimedia and Expo, pp. 549–552. IEEE,

    2006

  23. [2018]

    Instruction tuning with gpt-4

    Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277,

    arXiv

  24. [2019]

    The unreasonable effectiveness of eccentric automatic prompts

    Rick Battle and Teja Gollapudi. The unreasonable effectiveness of eccentric automatic prompts. arXiv preprint arXiv:2402.10949,

    arXiv

  25. [2020]

    Glide: Towards photorealistic image generation and editing with text-guided diffusion models

    Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741,

    arXiv

  26. [2021]

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

    Yirui Chen, Xudong Huang, Quan Zhang, Wei Li, Mingjian Zhu, Qiangyu Yan, Simiao Li, Hant- ing Chen, Hailin Hu, Jie Yang, et al. Gim: A million-scale benchmark for generative image manipulation detection and localization. arXiv preprint arXiv:2406.16531,

    arXiv

  27. [2022]

    Casia image tampering detection evaluation database

    11 Published as a conference paper at ICLR 2025 Jing Dong, Wei Wang, and Tieniu Tan. Casia image tampering detection evaluation database. In 2013 IEEE China summit and international conference on signal and information processing, pp. 422–426. IEEE,

    2025

  28. [2023]

    Instructdial: Improving zero and few-shot generalization in dialogue through instruction tuning

    Prakhar Gupta, Cathy Jiao, Yi-Ting Yeh, Shikib Mehri, Maxine Eskenazi, and Jeffrey P Bigham. Instructdial: Improving zero and few-shot generalization in dialogue through instruction tuning. arXiv preprint arXiv:2205.12673,

    arXiv

  29. [2024]

    Span: Spatial pyramid attention network for image manipulation localization

    Xuefeng Hu, Zhihan Zhang, Zhenye Jiang, Syomantak Chaudhuri, Zhenheng Yang, and Ram Nevatia. Span: Spatial pyramid attention network for image manipulation localization. In Computer Vision– ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part...

    2020

  30. [2025]

    Coverage—a novel database for copy-move forgery detection

    Bihan Wen, Ye Zhu, Ramanathan Subramanian, Tian-Tsong Ng, Xuanjing Shen, and Stefan Win- kler. Coverage—a novel database for copy-move forgery detection. In 2016 IEEE international conference on image processing (ICIP), pp. 161–165. IEEE,

    2016

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.