BG-REAL: A Public Real-Data Anchored Benchmark for Background Manipulation Detection and Localization

Bugra Alperen Uluirmak; Rifat Kurban

REVIEW 2 major objections 4 minor 24 references

BG-REAL: A Public Real-Data Anchored Benchmark for Background Manipulation Detection and Localization

T0 review · 2 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Background-manipulation benchmark surfaces a shared false-positive failure in forgery detectors.

desk verdict BG-REAL is a careful, honest benchmark for background manipulation, but the main empirical claim rests on a matched-control recipe the paper doesn't specify. read the letter →

arxiv 2607.26232 v1 pith:HZPT7S2R submitted 2026-07-28 cs.CV

Bugra Alperen Uluirmak , Rifat Kurban This is my paper

classification cs.CV

keywords imageforensicsmanipulationdetectionbackgroundbenchmarkmatchedcontrolsforgerylocalizationre-encodingartifactssource-disjointsplits

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper introduces BG-REAL, a publicly reproducible benchmark for detecting and localizing background manipulation in images. It is built from 7,000 samples anchored in real Open Images V7 photographs, spanning six edit types with matched authentic controls that re-encode untouched images through the same processing path as manipulated ones. The central empirical finding is that at a validation-fixed threshold, most blind detectors flag the majority of these re-encoded authentic images as manipulated—false-positive rates run from 0.57 to 1.00—so re-encoding artifacts are a shared shortcut risk rather than a quirk of one model. The package also provides source-disjoint splits, human-assisted quality control, five-seed evaluation, and official adapter baselines. If right, foreground-focused benchmarks have been missing a distinct evaluation axis, and BG-REAL offers a reproducible protocol for this axis.

What carries the argument

The matched-authentic-control edit family is the load-bearing construct: authentic images are passed through the same re-encoding and processing chain as manipulated samples, so any detector response to them can only come from processing artifacts, not manipulation content. Around this, the benchmark organizes a six-family taxonomy (authentic, matched authentic control, classic composite, harmonized composite, public background replacement, JPEG/resize robustness), source-group splits, mask and leakage QA, and an evaluation protocol with five seeds and validation-fixed thresholds. The pipeline's split audit distinguishes genuinely zero-leakage conditions (tool-OOD) from tag-filtered subset v

What would settle it

Run a detector on the matched-authentic-control split after replacing the pipeline's codec with a different one (e.g., PNG instead of JPEG); if the false-positive rate collapses, the shared shortcut is codec-specific rather than a general re-encoding effect. Alternatively, instrument the pipeline to confirm that controls and manipulated samples pass through identical re-encoding operators; any divergence invalidates the diagnostic.

Watch

Extended reading notes

Core claim

BG-REAL claims that background manipulation is a distinct, under-specified forensics setting and that a benchmark with matched authentic controls can measure it cleanly. On this benchmark, all three completed external baselines (TruFor, MVSS-Net, HiFi-Net) and the internally trained BG-RIFT model, when their thresholds are fixed on validation data, misclassify re-encoded authentic images as manipulated more than half the time; the rates are 0.57, 0.99, 1.00, and 0.64 respectively. The paper reads this as evidence that re-encoding-triggered false positives are a shared property of blind baselines, not a single-model failure. It also reports that source-disjoint splits produce only small AUROC

Load-bearing premise

The matched-authentic-control samples must genuinely undergo every non-manipulation step of the manipulated pipeline; if they skip or alter any step, the false-positive rates no longer isolate re-encoding artifacts.

Editorial extensions

If this is right

A detector can achieve strong AUROC on this benchmark while flagging most authentic re-encoded images, so accuracy tables alone are insufficient for deployment decisions.
TruFor, the strongest external baseline, still misclassifies 57% of matched authentic controls, suggesting even state-of-the-art forensic methods are not reliable on benign post-processing.
Localization behavior is split: on matched authentic controls TruFor predicts almost no affected pixels while HiFi-Net predicts large regions, so pooling localization across methods can obscure opposite failure modes.
The tool-OOD condition, though zero-leakage by construction, is synthetic-only and should be read as an infrastructure check, not evidence of real-world tool diversity.
Source-disjoint splits enable within-pipeline distribution-shift measurement, but the paper cautions that current background/generator OOD numbers are not generalization evidence.

Reading between the lines

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

The matched-authentic-control protocol could become a standard reporting item in other forensics benchmarks, because aggregate AUROC cannot reveal shortcut dependence on re-encoding.
If the proposed contrastive training signal is enabled in BG-RIFT, the matched-control false-positive rate may decrease measurably, offering a direct test of the paper's implied remedy.
A plausible next check is to re-run the same baselines on a conventional foreground-splicing benchmark with an identical matched-control protocol; the paper identifies this as future work, but it would test the motivating premise that background edits are harder than foreground edits.
The benchmark's design suggests a broader takeaway: real-world image forensics may need to separate 'manipulation content' from 'processing history' as two independent axes, rather than treating both as forgery.

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.

Desk Editor's Note

read the letter

Dear [Colleague],

BG-REAL is a solid pilot that does most things right. The idea of isolating background manipulation—away from foreground-object splicing—is a real gap, and the paper takes evaluation hygiene seriously: source-disjoint splits, a split-condition audit table (Table 8) that says exactly what each condition can and cannot support, five-seed summaries with bootstrap CIs and Holm-corrected paired tests, and external baselines run zero-shot through official adapters. The honest framing of what is not independent (tool-OOD as a leakage check, background/generator-OOD as tag-filtered subsets) is exactly what you want from a benchmark paper. The human-assisted QC is small (599 rows, single reviewer, AI pre-score) but the authors say so.

The soft spot is the matched-authentic-control construct, which is load-bearing for the paper's main empirical finding (§7: re-encoding-triggered false positives are widespread across blind baselines). The definition—'same processing path as manipulated samples'—is plausible but underspecified. The paper does not give a step-by-step recipe or pseudo-code for generating these controls, so I cannot tell whether a control for, say, a composite family goes through the same background resize/paste/final JPEG save as the manipulated sample, or is only a re-encode of the original. If the controls skip compositing-related operations, the high false-positive rates would not isolate re-encoding artifacts; they could reflect resolution or local-statistics mismatches. The stress-test note you passed along is right: this is unverified, and the code is not public until publication. Until that recipe is pinned down, the 'shared shortcut' claim is conditional.

Other caveats are minor for a pilot: the synthetic tool-OOD is explicitly not real-world evidence, the background/generator-OOD are subset views, and five seeds is a small sample—though the paper handles that gracefully with bootstrap intervals and a note that Wilcoxon can't reach significance at n=5. The provenance note about the relabeled diffusion_imported rows is slightly odd but transparent.

Bottom line: BG-REAL is a useful contribution for anyone building or evaluating background-manipulation detectors, provided the matched controls prove to be as matched as claimed. I would not desk-reject it; it deserves a serious referee, but the referee should ask for the exact construction steps (or a public, hashed code release) before the false-positive numbers can be taken at face value. If I were in the area, I'd cite it for the benchmark design, but I'd wait for the artifacts.

Best, [Your name]

Referee Report

2 major / 4 minor

Summary. BG-REAL is a benchmark package for background manipulation detection and localization, constructed from 1,000 Open Images V7 source images and expanded to 7,000 processed samples over 1,200 source groups. It defines six edit families — authentic, matched authentic control, classic composite, harmonized composite, public background replacement, and JPEG/resize robustness — and provides source-disjoint splits, human-assisted QA, three zero-shot official-adapter baselines (TruFor, MVSS-Net, HiFi-Net), and a weakly supervised internal model (BG-RIFT). The paper reports image-level and localization metrics on ID and source-OOD splits, plus a matched-authentic-control diagnostic in Section 7 that shows high false-positive rates across blind baselines, which the authors interpret as re-encoding artifacts being a shared shortcut risk. The authors are unusually careful in positioning the benchmark as a complement to general manipulation benchmarks and in disclosing that tool-OOD is a synthetic-only leakage check, background/generator OOD are tag-filtered subsets, and the foreground-vs-background difficulty premise is not directly tested.

Significance. If the matched-authentic-control protocol is exactly as claimed, the paper introduces a useful evaluation axis that most image-manipulation benchmarks lack, and the Section 7 diagnostic would be a valuable caution for the community. The paper also models good scientific hygiene: Table 8 explicitly audits what each split condition can and cannot support, Section 9 gives a candid limitations section, and the validation-fixed threshold protocol in Section 5 avoids a common source of optimistic bias. The main risk is that the paper's central empirical finding rests on the matched-control construction, whose exact recipe is currently underspecified, and the benchmark artifact itself is not yet publicly inspectable. With those gaps closed, I would regard this as a solid contribution.

major comments (2)

[Section 3, Table 2, and Section 7] The matched-authentic-control construct is load-bearing for the paper's main empirical finding, but the manuscript never specifies the processing recipe. Table 2 defines the family only as 're-encoded authentic samples that pass through the same processing path as manipulated samples,' and Section 7 refers to a 'processing chain' at the same level of abstraction. For each edit family (classic composite, harmonized composite, public background replacement), the paper should state exactly which operations the matched control undergoes: is the original background re-inserted after the same resize/crop/paste operations? Is the same JPEG quality factor and color-subsampling applied? Is the harmonization or blending step applied to the control? If controls differ from the manipulated condition in resolution, compression quality, or local statistics, then the high matched-control false-positive
[Section 10 and Section 7] The central deliverable is 'a reproducible public real-data anchored benchmark package,' but the exact artifact is not currently auditable. No repository URL, commit hash, or dataset DOI is given; the public code release and generated splits are 'finalized upon publication.' The matched-control false-positive rates in Section 7 and Figure 8 are produced by scripts that a reader cannot yet run or inspect, and the paper itself describes them as a single-snapshot diagnostic. Because the benchmark package is the scientific output, the split assignments, generation recipes, adapter code, and matched-control generation code should be available during review — for example, as a versioned repository or a supplementary archive with a checksum. Without this, the reproducibility claim cannot be verified, and the specific matched-control numbers in Section 7 remain uncheckable.

minor comments (4)

[Abstract and Section 7] The abstract quotes matched-control false-positive rates (0.57 to 1.00) without the paper's own caveat that these come from a single evaluation snapshot rather than the five-seed protocol. Please add the single-snapshot qualifier in the abstract or refer the reader to the caveat in Section 7.
[Section 6.1, Table 8 vs. Table 7] Table 8 lists n=295 for the matched-control condition, but each external baseline in Table 7 reports n=198 on that split while BG-RIFT reports n=295. The text explains that external adapters skip synthetic-control rows on the ID/source-OOD splits but not on the matched-control split. Please clarify whether the 97-sample difference is the synthetic-control subset, since the Section 7 false-positive rates for the external baselines are computed on 198 samples.
[Section 5, Table 6] For the RGB and Artifact baselines, the explanation that their validation-fixed thresholds are low enough to flag 'almost everything' would be easier to verify if the actual thresholds were listed in Table 6 or a footnote. This is a presentation improvement, not a substantive issue.
[Section 9] The sentence 'Split independence is now partial rather than fully absent' is awkward and could be misread as a protocol change. Rephrase as 'Split independence is partial: ...' and then give the breakdown already presented in Section 6.1.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's results are empirical measurements on external zero-shot baselines and a standard trained internal model, with disclosed limitations rather than definitional reductions.

full rationale

BG-REAL is a dataset/benchmark construction paper; it contains no formal derivation chain whose conclusions could reduce to its inputs. The main empirical result (Section 7) is a measurement: TruFor, MVSS-Net, and HiFi-Net run zero-shot from official weights, and BG-RIFT is trained on the benchmark's train split and evaluated on source-disjoint held-out splits using a decision threshold fixed once on validation. This is standard model evaluation, not a fitted parameter renamed as a prediction. The matched-authentic-control split is defined in Table 2 as 'Re-encoded authentic samples that pass through the same processing path as manipulated samples, isolating processing artifacts from manipulation evidence'; computing false-positive rates on it is a direct empirical probe of baseline behavior, not a conclusion forced by the definition. The paper explicitly cautions that the specific TruFor rate is threshold-sensitive and should be read as indicative, so no hidden forced conclusion is present. There are no self-citations used as load-bearing evidence, no imported uniqueness theorem, and no ansatz smuggled in via citation. The tag-filtered OOD conditions, synthetic controls, single-reviewer AI-assisted QC, and unmatched-control recipe concerns are disclosed in Sections 4, 6.1, 9, and 10; these are external-validity and reproducibility limitations, not circular reductions. One internal inconsistency exists — Table 7 reports non-zero IoU/pixel-F1 on the matched-control split whose affected masks are defined as empty, and Section 6.1 says BG-RIFT 'localizes the recorded intent region' — but this is a metric/pipeline consistency issue, not circularity in the paper's central claims.

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

The paper's central empirical contribution rests on two unvalidated-to-outsider data-construction assumptions: Open Images masks define valid manipulation contracts, and re-encoded matched controls isolate re-encoding artifacts. It also depends on threshold choices for both detection and localization metrics. All are disclosed in the text but not independently verifiable until the package is public.

free parameters (2)

Validation-fixed decision threshold (per baseline) = not reported numerically; chosen to maximize validation F1
F1/accuracy in Table 6 and matched-control false-positive rates in Fig. 8/§7 depend on thresholds selected once on the validation split; authors note TruFor's rate is sensitive to small threshold shifts.
Localization binarization threshold = 0.5
Used to binarize prediction heat maps for IoU and pixel F1 in Table 7; chosen by hand and shared across methods. Not fitted but influences all localization numbers.

assumptions (4)

domain assumption Open Images V7 instance-segmentation masks provide sufficiently accurate foreground/background separation to define intent masks for background editing.
The entire pipeline in §3 relies on Open Images V7 masks to mark foreground/background contracts; if masks are coarse or wrong, the generated manipulations and ground-truth affected masks inherit the error. The paper's own QA found 496 mask flags, retained rather than excluded.
domain assumption Re-encoding an authentic image through the same processing path as manipulated samples produces a matched control that differs only in the absence of manipulation.
Central to §7 matched-control FPR diagnostic (Table 2, §3). If the control path differs in other ways (e.g., compositing stages not applied), the false-positive rates do not isolate re-encoding artifacts.
domain assumption Source-group splits prevent leakage between train/validation/test.
Paper defines splits over source groups and reports duplicate SHA=0 and coarse perceptual-hash audits (§4), but cannot prove semantic leakage across images; 116 coarse perceptual-hash groups cross train/val and are not treated as duplicates.
domain assumption Official external-baseline adapters (TruFor, MVSS-Net, HiFi-Net) faithfully reproduce official outputs without ground-truth contamination.
The adapter hard-fails on missing masks and normalizes outputs (§5), but correctness depends on the authors' code, which is not yet public.

how reviews work

0 comments

Cite this review

Pith. "Pith review of BG-REAL: A Public Real-Data Anchored Benchmark for Background Manipulation Detection and Localization." pith.science (2026). https://pith.science/paper/HZPT7S2R

@misc{pith2026260726232,
  author       = {Pith},
  title        = {Pith review of: BG-REAL: A Public Real-Data Anchored Benchmark for Background Manipulation Detection and Localization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HZPT7S2R}},
  note         = {Machine review of arXiv:2607.26232}
}

read the original abstract

Background manipulation is a practical but under-specified image-forensics setting: the manipulated evidence can sit outside the salient foreground object, while many evaluations emphasize object-centric copy-move, splicing, or generic synthetic edits. We introduce BG-REAL, a public real-data anchored benchmark package for background manipulation detection and localization. The current release is built from Open Images V7 instance-segmentation sources and contains 7,000 processed samples over 1,200 source groups, including 6,000 public-data anchored samples and 1,000 synthetic control samples. BG-REAL covers six edit families, matched authentic controls, source-group splits, mask and leakage QA, 599 human-assisted quality-control rows, three completed external baselines (TruFor, MVSS-Net, and HiFi-Net), and five-seed model evaluation. Beyond aggregate accuracy, we use matched-authentic-control diagnostics to measure how often baselines misclassify re-encoded authentic images as manipulated at a threshold fixed on held-out validation data; false-positive rates range from 0.57 (TruFor, the lowest) to 1.00 (several weak or mask-informed baselines), indicating that re-encoding artifacts are a shared shortcut risk across baselines rather than a problem specific to any one model. The release provides the construction pipeline, evaluation protocol, paper-ready figures, and reproduction documentation. We frame BG-REAL as a background-manipulation-focused complement to general image-manipulation-localization benchmarks, not as a fully real-only or general-purpose benchmark.

Figures

Figures reproduced from arXiv: 2607.26232 by the authors.

**Figure 2.** Example BG-REAL edit family panels for two distinct public source groups (a food [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Dataset composition by edit family and source family. The public background re [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: Human-assisted quality-control score distributions for the reviewed subset. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: Split allocation and QA audit summary. 9 [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

**Figure 6.** Figure 6: Five-seed baseline performance on ID and source-OOD splits. The AUROC panels [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

**Figure 7.** Figure 7: Localization quality (mean pixel F1, mean IoU), pooled across TruFor, MVSS-Net, [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]

**Figure 8.** Figure 8: Left: false-positive rate (1−accuracy) on the matched-authentic-control split, where every sample is authentic but has been re-encoded through the same processing chain as a manipulated sample. Right: AUROC on the in-distribution test split versus the disjoint samepip…

**Figure 9.** Figure 9: Qualitative localization examples across three distinct source images and edit fam [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 6 canonical work pages

[1]

Kuznetsova, H

A. Kuznetsova, H. Rom, N. Alldrin, J. Uijlings, I. Krasin, J. Pont-Tuset, S. Kamali, S. Popov, M. Malloci, A. Kolesnikov, T. Duerig, V. Ferrari, The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale, International Journal of Computer Vision (2020).doi:10.1007/s11263-020-01316-z

doi:10.1007/s11263-020-01316-z 2020
[2]

Google Open Images, Open images v7 facts and figures,https:// storage.googleapis.com/openimages/web/factsfigures_v7.html, ac- cessed 2026-07-16 (2022). 21

2026
[3]

S. Xue, A. Agarwala, J. Dorsey, H. E. Rushmeier, Understanding and im- proving the realism of image composites, ACM Transactions on Graphics 31 (4) (2012).doi:10.1145/2185520.2185580

arXiv 2012
[4]

J.-Y. Zhu, P. Krahenbuhl, E. Shechtman, A. A. Efros, Learning a dis- criminative model for the perception of realism in composite images, in: Proceedings of the IEEE International Conference on Computer Vision, 2015.doi:10.1109/iccv.2015.449

work page doi:10.1109/iccv.2015.449 2015
[5]

W.Cong, J.Zhang, L.Niu, L.Liu, Z.Ling, W.Li, L.Zhang, Dovenet: Deep image harmonization via domain verification (2020).arXiv:1911.13239, doi:10.48550/arxiv.1911.13239

work page Pith review arXiv doi:10.48550/arxiv.1911.13239 2020
[6]

L. Niu, W. Cong, L. Liu, Y. Hong, B. Zhang, J. Liang, L. Zhang, Making images real again: A comprehensive survey on deep image composition (2021).arXiv:2106.14490,doi:10.48550/arxiv.2106.14490

doi:10.48550/arxiv.2106.14490 2021
[7]

Mareen, D

H. Mareen, D. Karageorgiou, G. Van Wallendael, P. Lambert, S. Pa- padopoulos, Tgif: Text-guided inpainting forgery dataset, in: IEEE In- ternational Workshop on Information Forensics and Security, 2024.doi: 10.1109/wifs61860.2024.10810690

arXiv 2024
[8]

Mareen, D

H. Mareen, D. Karageorgiou, P. Giakoumoglou, P. Lambert, S. Papadopou- los, G. Van Wallendael, Tgif2: Extended text-guided inpainting forgery dataset & benchmark, accepted at Journal on Information Security (2026). arXiv:2603.28613,doi:10.48550/arxiv.2603.28613

work page doi:10.48550/arxiv.2603.28613 2026

Show all 24 references

[10]

C. Dong, X. Chen, R. Hu, J. Cao, X. Li, Mvss-net: Multi-view multi-scale supervised networks for image manipulation detection, IEEE Transactions onPatternAnalysisandMachineIntelligence(2022).doi:10.1109/TPAMI. 2022.3180556

2022
[11]

X. Guo, X. Liu, Z. Ren, S. Grosz, I. Masi, X. Liu, Hierarchical fine-grained image forgery detection and localization, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023.doi:10. 1109/CVPR52729.2023.00308

2023
[12]

J. Dong, W. Wang, T. Tan, Casia image tampering detection evalua- tion database, in: IEEE China Summit and International Conference on Signal and Information Processing, 2013.doi:10.1109/chinasip.2013. 6625374

2013 doi
[13]

B. Wen, Y. Zhu, R. Subramanian, T.-T. Ng, X. Shen, S. Winkler, Coverage—a novel database for copy-move forgery detection, in: IEEE 22 International Conference on Image Processing, 2016.doi:10.1109/icip. 2016.7532339

2016
[14]

Novozamsky, B

A. Novozamsky, B. Mahdian, S. Saic, Imd2020: A large-scale anno- tated dataset tailored for detecting manipulated images, in: IEEE Win- ter Conference on Applications of Computer Vision Workshops, 2020. doi:10.1109/wacvw50321.2020.9096940

2020
[15]

Mahfoudi, B

G. Mahfoudi, B. Tajini, F. Retraint, F. Morain-Nicolier, J.-L. Dugelay, M. Pic, Defacto: Image and face manipulation dataset, in: Proceed- ings of the European Signal Processing Conference, 2019.doi:10.23919/ EUSIPCO.2019.8903181

2019
[16]

github.io/, accessed 2026-07-16 (2019)

DEFACTO Dataset, Defacto dataset website,https://defactodataset. github.io/, accessed 2026-07-16 (2019)

2026
[17]

GRIP, University of Naples Federico II, Trufor official repository and project page,https://grip-unina.github.io/TruFor/, accessed 2026- 07-16 (2023)

2026
[18]

Dong, Mvss-net official repository,https://github.com/dong03/ MVSS-Net, accessed 2026-07-16 (2022)

C. Dong, Mvss-net official repository,https://github.com/dong03/ MVSS-Net, accessed 2026-07-16 (2022)

2026
[19]

X. Guo, X. Liu, Z. Ren, S. Grosz, I. Masi, X. Liu, Hifi-ifdl official repos- itory,https://github.com/chelsea234/hifi_ifdl, accessed 2026-07-16 (2023)

2026
[20]

J. Wang, Z. Wu, J. Chen, X. Han, A. Shrivastava, S.-N. Lim, Y.-G. Jiang, Objectformer for image manipulation detection and localization, in: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022.doi:10.1109/cvpr52688.2022.00240

2022
[21]

Kwon, S.-H

M.-J. Kwon, S.-H. Nam, I.-J. Yu, H.-K. Lee, C. Kim, Learning jpeg com- pression artifacts for image manipulation detection and localization (2021). arXiv:2108.12947,doi:10.48550/arxiv.2108.12947

2021 doi arXiv
[22]

Zhang, M

Z. Zhang, M. Li, M.-C. Chang, A new benchmark and model for challenging image manipulation detection, Proceedings of the AAAI Conference on Artificial Intelligence (2024).doi:10.1609/aaai.v38i7.28571

2024 doi
[23]

Rössler, D

A. Rössler, D. Cozzolino, L. Verdoliva, C. Riess, J. Thies, M. Nießner, Faceforensics++: Learning to detect manipulated facial images, in: Pro- ceedings of the IEEE/CVF International Conference on Computer Vision, 2019.doi:10.1109/iccv.2019.00009

2019
[24]

X. Ma, X. Zhu, L. Su, B. Du, Z. Jiang, B. Tong, Z. Lei, X. Yang, C.-M. Pun, J. Lv, et al., Imdl-benco: A comprehensive benchmark and code- base for image manipulation detection & localization, Advances in Neural Information Processing Systems 37 (2025) 134591–134613. 23

2025
[25]

Nixon, M

J. Nixon, M. Dusenberry, G. Jerfel, T. Nguyen, J. Liu, L. Zhang, D. Tran, Measuring calibration in deep learning (2019).arXiv:1904.01685,doi: 10.48550/arxiv.1904.01685. 24

2019 doi arXiv

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

Kuznetsova, H

A. Kuznetsova, H. Rom, N. Alldrin, J. Uijlings, I. Krasin, J. Pont-Tuset, S. Kamali, S. Popov, M. Malloci, A. Kolesnikov, T. Duerig, V. Ferrari, The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale, International Journal of Computer Vision (2020).doi:10.1007/s11263-020-01316-z

doi:10.1007/s11263-020-01316-z 2020

[2] [2]

Google Open Images, Open images v7 facts and figures,https:// storage.googleapis.com/openimages/web/factsfigures_v7.html, ac- cessed 2026-07-16 (2022). 21

2026

[3] [3]

S. Xue, A. Agarwala, J. Dorsey, H. E. Rushmeier, Understanding and im- proving the realism of image composites, ACM Transactions on Graphics 31 (4) (2012).doi:10.1145/2185520.2185580

arXiv 2012

[4] [4]

J.-Y. Zhu, P. Krahenbuhl, E. Shechtman, A. A. Efros, Learning a dis- criminative model for the perception of realism in composite images, in: Proceedings of the IEEE International Conference on Computer Vision, 2015.doi:10.1109/iccv.2015.449

work page doi:10.1109/iccv.2015.449 2015

[5] [5]

W.Cong, J.Zhang, L.Niu, L.Liu, Z.Ling, W.Li, L.Zhang, Dovenet: Deep image harmonization via domain verification (2020).arXiv:1911.13239, doi:10.48550/arxiv.1911.13239

work page Pith review arXiv doi:10.48550/arxiv.1911.13239 2020

[6] [6]

L. Niu, W. Cong, L. Liu, Y. Hong, B. Zhang, J. Liang, L. Zhang, Making images real again: A comprehensive survey on deep image composition (2021).arXiv:2106.14490,doi:10.48550/arxiv.2106.14490

doi:10.48550/arxiv.2106.14490 2021

[7] [7]

Mareen, D

H. Mareen, D. Karageorgiou, G. Van Wallendael, P. Lambert, S. Pa- padopoulos, Tgif: Text-guided inpainting forgery dataset, in: IEEE In- ternational Workshop on Information Forensics and Security, 2024.doi: 10.1109/wifs61860.2024.10810690

arXiv 2024

[8] [8]

Mareen, D

H. Mareen, D. Karageorgiou, P. Giakoumoglou, P. Lambert, S. Papadopou- los, G. Van Wallendael, Tgif2: Extended text-guided inpainting forgery dataset & benchmark, accepted at Journal on Information Security (2026). arXiv:2603.28613,doi:10.48550/arxiv.2603.28613

work page doi:10.48550/arxiv.2603.28613 2026