acceptodds
Under review as a conference paper at ICLR 2027

Impostor: An Agent-Curated Benchmark for Realistic AIGC Manipulation Localization

Abstract

Recent advances in generative image editing have improved the realism and controllability of localized manipulation, raising new challenges for image manipulation detection and localization (IMDL). However, existing IMDL benchmarks have limitations in visual realism, manipulation diversity, and generator coverage, making it difficult to reflect recent trends in image manipulation. To address these limitations, we introduce Impostor, a high-quality AI-edited image manipulation localization dataset containing 100K manipulated images. Impostor is constructed by Disguise-Agent (DisAgent), a closed-loop agent framework that integrates scene perception, editing planning, manipulation execution, quality validation, and iterative reflection to automatically generate diverse and visually realistic manipulated images together with pixel-level masks, without requiring manual annotation. Moreover, Impostor contains images generated by seven recent AIGC models across three manipulation types and includes multiple manipulated regions, providing a more comprehensive benchmark for AIGC-based IMDL. Furthermore, we propose PhaseAware-Net (PANet), a semantic-forensic framework that introduces local phase modeling and Forensic-Semantic Consistency (FSC) learning to better localize semantically plausible yet forensically disrupted manipulated regions. Extensive experiments show that Impostor poses significant challenges to existing large vision-language models (LVLMs) and specialized IMDL methods, while PANet achieves superior performance on Impostor and multiple public benchmarks.

Primary area:
datasets and benchmarks
Keywords:
Image manipulation detection and localization (IMDL), AI-generated Image Manipulation Benchmark

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_gof6OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Impostor, an agent-curated benchmark for realistic AIGC manipulation localization, and proposes PANet, a semantic-forensic framework for detecting and localizing AI-edited image regions. The paper identifies limitations in existing AIGC manipulation localization datasets, including insufficient visual realism, limited manipulation diversity, and narrow generator coverage. To address these issues, the authors propose DisAgent, a closed-loop agent framework that performs perception, planning, execution, validation, and reflection to automatically generate realistic manipulated images with pixel-level masks. The resulting Impostor dataset contains 100K manipulated image pairs generated from seven recent AIGC editing models. Furthermore, the authors propose PANet, which combines local forensic cue extraction with forensic-semantic consistency learning to identify subtle manipulation traces. Extensive experiments show strong localization performance and cross-domain generalization.

Critical strengths
  1. Significant benchmark contribution. The paper addresses an important limitation in current AIGC manipulation localization research: existing benchmarks often fail to reflect the realism and diversity of modern AI editing systems. The proposed Impostor dataset is a valuable contribution because it provides large-scale manipulated image pairs with pixel-level masks and covers multiple manipulation types and generators. The benchmark design is likely to be useful for future research in image forensics.
  2. Agent-based dataset construction is well motivated. The use of DisAgent to simulate a realistic editing workflow is an interesting idea. Instead of relying on fixed prompts, the agent considers scene context, editing feasibility, and iterative quality feedback, which helps generate more realistic manipulation samples. The closed-loop validation and reflection mechanism is a meaningful improvement over simple sequential generation pipelines.
  3. The proposed PANet framework is technically reasonable. The semantic-forensic consistency perspective provides a useful insight for AIGC manipulation localization. Modeling the mismatch between semantic plausibility and forensic traces is appropriate for modern AI-edited images, where visual boundaries can appear highly realistic.
  4. Extensive experimental evaluation. The paper provides comprehensive experiments including in-domain evaluation, cross-domain generalization, ablation studies, and robustness analysis. The results demonstrate consistent improvements over existing methods.
Critical weaknesses
  1. The benchmark contribution dominates the algorithmic contribution. The main contribution of the paper is clearly the Impostor dataset. While PANet introduces useful modules such as local phase modeling and FSC learning, the methodological novelty is relatively incremental compared with the benchmark contribution. The proposed architecture still follows the common pattern of combining semantic features and forensic features with additional supervision losses. The authors should better clarify the fundamental algorithmic insight beyond the specific design choices.
  2. The realism evaluation may not be fully sufficient. The paper uses human verification and LVLM-based evaluation to measure realism. However, realism is inherently difficult to quantify, and human judgments may still be influenced by image content or visual preferences. More rigorous analysis of whether the benchmark truly reduces shortcut learning would be helpful.
  3. The DisAgent contribution is difficult to separate from engineering components. The improvement from the agent-based construction pipeline is interesting, but the paper does not fully analyze which component contributes most. It remains unclear how much benefit comes from agent reasoning itself versus the use of stronger editing models and filtering criteria.
  4. Limited discussion of dataset release risks. The paper plans to release realistic AI manipulation data, masks, prompts, and generation metadata. Although this is valuable for research, such resources could potentially facilitate creation of deceptive images. The authors briefly acknowledge this issue, but a more detailed discussion of mitigation strategies would strengthen the responsible research aspect.
Questions for authors
  1. How much of the improvement in PANet comes from FSC learning compared with the local forensic feature extraction module?
  2. How well does PANet generalize to real-world manipulated images created by human editors rather than AIGC models?
  3. Could DisAgent be applied to continuously expand the benchmark as new image editing models emerge?
  4. What mechanisms will be used to mitigate potential misuse of the released manipulation dataset?
4 comments
AuthorsOpenReview ↗

Response Part 1

Dear Reviewer gof6,

We sincerely thank you for your thoughtful and constructive feedback. We appreciate your recognition of Impostor’s benchmark contribution, the DisAgent design, and PANet’s methodology. Below, we address your concerns and questions point by point.

W1 & Q1: PANet’s Core Methodological Insight and Component Contributions

Thank you for your insightful analysis of PANet’s design.

First, regarding PANet’s core methodological insight, we would like to clarify that although PANet adopts a dual-stream architecture, its key idea is not simply to fuse features or stack auxiliary losses. Instead, PANet jointly models local phase structures and forensic–semantic consistency as complementary cues for manipulation localization. Modern AIGC editing can generate semantically plausible and visually natural content while disrupting local forensic traces and weakening their consistency with semantic representations. Motivated by this observation, PANet is designed around two complementary aspects:

  • Local Phase Structure Modeling: As analyzed in Appendix A.2, the LPCE block models spatially aligned local phase patterns to capture subtle structural disruptions in edges, contours, and textures that conventional frequency-domain statistics may overlook. This offers a new perspective on image manipulation localization by leveraging local phase information.

  • Forensic–Semantic Consistency Modeling: FSC exploits the observation that AIGC editing can weaken the correspondence between semantic representations and local forensic cues. It explicitly supervises the relative consistency gap between authentic and manipulated regions, turning this difference into discriminative guidance for manipulation localization.

These design choices are supported by independent empirical evidence rather than being based solely on heuristic module combinations. The diagnostic experiment in Appendix A.1 supports the hypothesis that AIGC editing weakens local forensic–semantic consistency. In addition, the comparison with DCT-, DWT-, and FFT-based representations in Appendix A.2 further demonstrates the discriminative ability of local phase responses in distinguishing manipulated regions.

Second, to quantify the contributions of LFCE and FSC, we report the corresponding ablation results in Table 4, as summarized below.

VariantAverage IoU ↑\uparrowPerformance Drop (pp)
Full PANet59.78—
w/o LFCE (Raw RGB)56.043.74
w/o LPCE (RCE-only)58.341.44
w/o RCE (LPCE-only)58.171.61
w/o FSC58.371.41

Table A: Contributions of forensic representation and FSC learning (Average IoU, %).

As shown in Table A, removing LFCE reduces the average IoU by 3.74%. Removing FSC reduces the average IoU by 1.41%. These results indicate that LFCE has a larger impact on localization performance, while FSC provides an additional gain through explicit forensic–semantic consistency supervision. Overall, LFCE and FSC play distinct yet complementary roles in PANet: LFCE enhances local forensic cue extraction, while FSC exploits forensic–semantic consistency to guide manipulation localization.

To further illustrate the role of FSC, we have added Figure 4 on page 10 of the revised manuscript, comparing localization results with and without FSC. The visualization shows how FSC turns local forensic–semantic inconsistencies into discriminative localization cues, reducing missed detections within manipulated regions and improving localization completeness.

AuthorsOpenReview ↗

Response Part 2

W2: Reliability of Visual Realism Evaluation

Thank you for your valuable suggestions regarding the visual realism evaluation of Impostor.

We agree that visual realism cannot be fully and objectively quantified by a single metric. To provide a more systematic and comprehensive assessment, we define visual realism from a perceptual perspective. Specifically, a manipulated image is considered visually realistic if it exhibits perceptual consistency with naturally captured photographs in terms of scene structure, lighting, textures, and local details. Based on this definition, we combine human verification with evaluations from multiple large vision-language models (LVLMs) to provide complementary assessments rather than relying on the judgment of a single LVLM.

Second, to further assess the reliability of LVLM-based evaluations, we randomly sampled 1,000 authentic images from Impostor and measured each LVLM’s accuracy in identifying them as authentic (Table B).

LVLMReal Image Recognition Accuracy (%) ↑\uparrow
Qwen3-VL-235B-A22B-Instruct92.5%
FakeShield94.1%
Gemini 3 Pro93.2%
GPT-5.499.5%

Table B: Recognition accuracy of LVLMs on 1,000 authentic images.

As shown in Table B, the involved LVLMs all achieve high accuracy, demonstrating their ability to reliably distinguish authentic images. Together with the results in Table 1, this experiment provides two complementary validations of the reliability of LVLM-based evaluations. First, the LVLMs can accurately recognize authentic images. Second, their detection rates are substantially higher on other manipulated-image benchmarks than on Impostor, suggesting that the lower detection rate on Impostor cannot be simply attributed to evaluator bias. To further mitigate potential bias, we employ multiple general-purpose and forensic-specific LVLMs under a unified evaluation protocol. Furthermore, Appendix C.2 (Tables 11–13) examines the effects of manipulated-area ratios, manipulation types, and the number of manipulated regions on detection difficulty, indicating that the lower detection rate on Impostor cannot be explained solely by these dataset-composition factors. Combined with human quality verification, these findings provide consistent and complementary evidence that Impostor exhibits higher visual realism than the compared AIGC manipulation datasets.

Finally, to further examine whether Impostor's emphasis on visual realism helps models learn more transferable manipulation cues, we conducted a controlled experiment.

(1) Controlled Experiment on Visual Realism. To investigate the effect of visual realism on shortcut learning, we constructed two subsets of 20,000 manipulated images each from DisAgent's data construction process. The Low-Realism subset consists of samples rejected during automatic quality filtering or human naturalness review due to obvious visual artifacts, while the High-Realism subset consists of Impostor samples that passed the complete quality validation and human review processes. Both subsets were drawn from the same data sources and generator pool, with identical training sample sizes and configurations.

For each subset, we used the same set of 20,000 authentic images as negative samples, maintaining a 1:1 ratio between authentic and manipulated images. Under identical training configurations, we independently trained TruFor, NFA-ViT, and PANet on the two subsets and evaluated their image-level detection performance on three external test sets: AutoSplice, BR-Gen, and CocoGlide.

MethodTraining SubsetAutoSpliceBR-GenCocoGlide
TruForLow-Realism58.1758.2552.13
TruForHigh-Realism67.1667.0054.85
NFA-ViTLow-Realism70.4159.1554.13
NFA-ViTHigh-Realism78.9469.4059.24
PANetLow-Realism69.2164.4657.23
PANetHigh-Realism81.8273.2263.65

Table C: Effect of training data visual realism on cross-domain image-level detection (Balanced Accuracy, % ↑\uparrow).

As shown in Table C, training on the High-Realism subset consistently yields better image-level detection performance across all three methods and external test sets. In particular, PANet achieves absolute BA gains of 12.61%, 8.76%, and 6.42% on AutoSplice, BR-Gen, and CocoGlide, respectively.

These results suggest that, compared with low-realism images containing obvious visual artifacts, high-realism manipulated images help models learn more transferable manipulation cues and reduce reliance on such artifacts as shortcuts, thereby improving cross-domain generalization.

AuthorsOpenReview ↗

Response Part 3

Q2: Generalization to Human-Edited Manipulations

Thank you for your valuable suggestion regarding PANet's generalization to human-edited manipulations.

We agree that, although PANet is primarily designed for AIGC manipulation localization, evaluating its performance on conventional non-AIGC manipulations is also important for understanding its generalization across manipulation types. To this end, we conducted an additional zero-shot cross-domain evaluation on conventional image manipulation benchmarks.

Specifically, we directly evaluated the model checkpoints trained on Impostor in Sec. 5.2.1 on five conventional manipulation benchmarks: CASIA V1, CASIA V2, Columbia, Coverage, and IMD2020. All comparison methods followed the same evaluation protocol as in Sec. 5.2.1, with pixel-level localization performance measured only on manipulated images.

MethodCASIA V1(918)CASIA V2(3,000)Columbia(180)Coverage(91)IMD2020(2020)Average
MVSSNet14.649.6218.189.477.4611.87
TruFor15.0112.4611.7912.4312.3712.81
Mesorch29.1025.3219.269.9718.0520.34
M2SFormer30.0820.4022.3112.7319.3620.98
NFA-ViT23.0023.1321.9810.4416.9219.09
PSCC-Net14.9915.5012.9211.4917.7514.53
RITA29.8426.0725.5012.3921.7623.11
PANet31.9328.1418.7215.6023.6323.60

Table D: Zero-shot generalization from Impostor to conventional image manipulation benchmarks (IoU, % ↑\uparrow).

As shown in Table D, without additional training on conventional manipulation data, PANet achieves the best localization performance on four of the five benchmarks: CASIA V1, CASIA V2, Coverage, and IMD2020. Although PANet does not achieve the best result on Columbia, its average IoU across all five benchmarks reaches 23.60%, outperforming the strongest baseline, RITA (23.11%), with an absolute IoU gain of 0.49%.

These results suggest that PANet exhibits generalization capability across different manipulation types, highlighting its potential applicability to broader image forensics scenarios.

AuthorsOpenReview ↗

Response Part 4

W3: Clarifying the Core Contribution of DisAgent

Thank you for your interest in the contributions of DisAgent's components.

Appendix E.1 (Table 16) already includes ablation experiments on DisAgent. Specifically, Perception, Planning, and Execution constitute the basic data generation pipeline. We therefore keep these three stages fixed across all ablation settings, while independently controlling the two key mechanisms of DisAgent's design: Validation and Reflection.

To provide a clearer comparison of component contributions, we average the results across the four image-level detection models and four pixel-level localization models reported in Table 16, as summarized below.

SettingValidationReflectionImage Det. (%) ↓\downarrowPixel IoU (%) ↓\downarrow
Sequential Pipeline✗✗33.6714.56
w/o Validation✗✓29.6313.76
w/o Reflection✓✗21.7012.70
Full DisAgent✓✓19.3712.28

Table E: Contributions of Validation and Reflection in DisAgent.

The results show that Validation is the primary contributor to improving sample quality, while Reflection provides additional improvements both with and without Validation. Notably, even without quality filtering, Reflection can use historical sample statistics to guide subsequent planning and improve generation outcomes. This demonstrates that feedback-driven iterative decision-making contributes independently of quality filtering.

Second, to examine the influence of generator capabilities, we compare different DisAgent configurations across several generators.

GeneratorSequentialValidation OnlyReflection OnlyFull DisAgent
Nano Banana Pro23.6319.1122.7614.08
Step1X-Edit33.7523.3331.4020.81
FLUX.1 Kontext39.7427.0434.8424.62
Qwen-Image-Edit45.9330.7839.8325.87

Table F: Generator-specific ablation of DisAgent under Add operations (Image Det., % ↓).

The results show that generator capabilities do affect sample detectability, with stronger generators generally providing a better starting point for generating realistic manipulations. However, with the generator and editing operation fixed, both Validation and Reflection consistently improve the results, and their combination further increases the detection difficulty of the generated samples.

This indicates that DisAgent's improvements do not rely solely on stronger generators; rather, its closed-loop mechanisms contribute consistently across different editing models.

 

Q3: Extensibility of DisAgent to New Editing Models

Thank you for your interest in DisAgent’s extensibility. DisAgent is not tied to any specific image editing model. New open- or closed-source models can be integrated into the Execution stage through local deployment or API calls. They can then be used within the existing Perception, Planning, Validation, and Reflection framework. The generated samples follow the same automated quality validation and feedback process, without requiring any redesign of the framework. As described in Sec. 3.2 (Lines 185–194), this modular design allows DisAgent to continuously expand Impostor with emerging AIGC editing models. We have also integrated GPT-Image-2 through this pipeline and generated additional test samples, which we plan to release together with Impostor.

 

W4 & Q4: Responsible Dataset Release and Misuse Mitigation

Thank you for highlighting the dataset release risks. We recognize that releasing realistic AI manipulation data, masks, prompts, and generation metadata could potentially facilitate the creation of deceptive images.

To mitigate these risks, we plan to adopt a tiered release strategy. We will make the manipulated images, masks, dataset splits, and evaluation code publicly available to support IMDL research. Access to detailed editing prompts, generation parameters, and code for reproducing the editing pipeline will be restricted to approved research requests and subject to usage agreements. These agreements will specify permitted research uses and explicitly prohibit using the released resources to create deceptive images or conduct other malicious activities. We will also review applicable data licenses and screen for potentially sensitive content before release.

We have expanded the revised Ethics Statement to include these concrete mitigation strategies, supporting responsible research and reproducibility while reducing the potential for misuse.

Reviewer_cuLGOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper builds Impostor, a localization benchmark of 100,000 source-manipulated pairs with masks and generation metadata, and PANet, a two-branch localizer trained on it. DisAgent is a closed loop of perception, planning, generator execution, four-way validation, and reflection. Seven editors are used (FLUX.1 Fill, FLUX.1 Kontext, PowerPaint, Step1X-Edit, Qwen-Image-Edit, OmniEraser, and closed-source Nano Banana Pro), with add, delete, and inpaint and up to nine regions per image. Two annotators then keep a sample only if both accept it as a natural photograph and accept the mask; this removes about 16% of validated candidates. Each generator is split 7:1, and a source image stays with all of its edits. PANet uses SegFormer-B2 for semantics and ConvNeXt V2-Tiny for forensics (74.9M parameters). The forensic stream combines constrained residual filters with a local-phase block (four Log-Gabor scales), fused by mutual guided cross-attention. Training adds a forensic-semantic consistency (FSC) loss and a TraceNCE loss to Dice-BCE segmentation and image-level BCE. Trained on Impostor, PANet reports average localization IoU/F1 of 59.78/66.03 across Impostor, SID-Set, BR-Gen, CocoGlide, and AutoSplice, 5.15 IoU above the next best average (RITA, 54.63). In-domain Impostor IoU is 78.59. A separate ProGAN-to-many-generators detection setup, using two 32x32 DCT patches, reports mean accuracy 94.42 versus AIDE 91.25, and 71.19 on Impostor versus 67.01.

Critical strengths

First, Impostor is hard for models trained on the recent AIGC sets the paper compares against, and the comparison is not limited to PANet. Table 7 shows double-digit IoU on Impostor after training on BR-Gen or SID-Set, for PANet, Mesorch, and M2SFormer, while training on Impostor transfers better for all three. Table 9 also shows that difficulty is not uniform: FLUX Kontext addition reaches 87.99 IoU for PANet, while Nano Banana addition is 57.31 and OmniEraser deletion is 60.83. A benchmark with that spread, at this scale, is consequential for the area. Second, the paper does not only assert realism and diversity. It reports a human reject rate of about 16%, a source-level split that blocks trivial pair leakage, seven named editors, and a training-source swap that holds for more than one architecture. Those are the facts I would use in an accept case.

Critical weaknesses

First, the official split cannot support the generalization claim as written. Section 3.2 argues that few-generator sets invite generator shortcuts, but Section 3.4 puts every generator in train and test. There is no leave-one-generator-out result. Cross-domain IoU is high on CocoGlide and AutoSplice (66.36 and 74.82) and much lower on BR-Gen (32.69). "Strong cross-domain generalization" in Section 5.2.1 averages over that gap. I would not accept the paper on that sentence. Second, the method claim is ahead of the ablation. Removing FSC costs 1.41 IoU and removing TraceNCE costs 0.58, yet Section 5.2.1 attributes the gains to forensic-semantic mismatch. No paired source-versus-edit oracle is reported, even though validation itself is defined around agreement inside and outside the mask, and I could not find a statement of whether the source image is withheld at test time. If a feature-difference baseline is already near the in-domain 78.59 IoU, PANet is not what solves the task. The two Impostor in-domain numbers, 78.59 and 80.12, also need a protocol explanation before the leaderboard can be checked.

Questions for authors

Clarification: At test time, is the unedited source image available to the localizer? What IoU does a pixel-difference or frozen-feature difference between source and edit obtain on the Impostor test split? Clarification: Please reconcile PANet's Impostor IoU of 78.59 (Table 2), 80.12 (Table 8), and the subset average 76.96 (Table 9). Which protocol difference produces the two-point shift between Tables 2 and 8, which appears for the other methods as well? Clarification: Is there a leave-one-generator-out evaluation, and which generators in BR-Gen overlap FLUX or PowerPaint? The answer would change how much of Table 2 I read as transfer. These are clarification requests. A human audit of addition masks, which are produced after generation rather than taken from the editor, would also affect the soundness score, but I am not asking for a new research program beyond that check.

4 comments
AuthorsOpenReview ↗

Response Part 1

Dear Reviewer cuLG,

Thank you for your valuable feedback. We appreciate your recognition of Impostor’s benchmark value and our data quality controls. Below, we address your concerns and questions point by point.

W2 & Q1: Test-Time Source Image Availability and Difference-Based Baselines

Thank you for your careful review. Below, we address your questions about PANet’s design and test-time evaluation protocol.

First, regarding the attribution of performance gains, we agree that attributing the overall improvement primarily to forensic–semantic mismatch in Sec. 5.2.1 was too strong and did not fully reflect the contributions of different components. In fact, PANet’s performance gains come from jointly modeling local phase structures and forensic–semantic consistency, rather than relying on a single mechanism, as supported by the ablation study in Sec. 5.3. We have revised this statement in the revised manuscript to provide a more balanced attribution.

Second, we clarify that the IMDL task studied in this work follows a single-image setting: models take only the image under examination as input, without access to its corresponding source image at test time. A trustworthy source image is often unavailable in practical image forensics. Impostor provides paired source and manipulated images primarily for dataset construction, annotation quality verification, and further research. MA-BC (Appendix E.5) uses paired feature comparisons inside and outside the mask exclusively for data quality filtering.

Furthermore, following your suggestion, we conducted localization experiments on the Impostor test set using feature-difference baselines. Specifically, Pixel Difference computes absolute pixel-wise differences between the source and manipulated images after normalizing their pixel values to [0, 1]. For Frozen-Feature Difference, we evaluate two variants based on frozen DINOv2 and ResNet encoders. Each variant applies the same encoder to both images, computes cosine similarities between spatially corresponding features, and converts them into normalized difference maps. For each baseline, we apply different thresholds to the difference maps to obtain binary predicted masks. Importantly, these baselines use the source image paired with each manipulated image at test time, providing privileged reference information unavailable in the single-image IMDL setting. This allows them to localize manipulations directly from differences between the paired images. We therefore report them as reference-based comparisons rather than direct competitors to PANet.

SettingMethodInputThresholdsIoU
Single-image IMDLPANetEdited image only—78.59
Reference-basedPixel DifferenceSource + Edited0, 0.05, 0.1, 0.2, 0.310.07 / 56.09 / 57.49 / 41.87 / 28.01
Reference-basedFrozen-Feature Difference (DINOv2)Source + Edited0, 0.05, 0.1, 0.2, 0.39.70 / 52.25 / 54.05 / 32.60 / 3.60
Reference-basedFrozen-Feature Difference (ResNet)Source + Edited0, 0.05, 0.1, 0.2, 0.39.70 / 22.29 / 35.28 / 57.62 / 52.59

Table A: Localization performance under different input settings on Impostor.

As shown in Table A, even with paired source images at test time, difference-based methods underperform PANet, which uses only the manipulated image. Across the tested thresholds, Pixel Difference, DINOv2-based feature difference, and ResNet-based feature difference achieve their highest IoU scores of 57.49%, 54.05%, and 57.62%, respectively, all below PANet's 78.59%.

Simple differencing methods measure the magnitude of changes in pixel values or frozen features, whereas the GT masks represent the target manipulated objects or designated editing regions, rather than all locations where pixel values change. Since generative editing may introduce minor reconstruction differences in non-target regions, difference-based methods may incorrectly identify these regions as manipulated while missing parts of the target region where changes are subtle. Furthermore, Table A shows that their localization performance is sensitive to threshold selection, with substantial variations across different thresholds. In contrast, PANet does not require access to the source image and achieves more accurate localization.

AuthorsOpenReview ↗

Response Part 2

Q2: Clarification of the Different Impostor IoU Results

Thank you for carefully checking these experimental results. The IoU differences across Tables 2, 8, and 9 arise from different evaluation objectives, checkpoint selection criteria, and aggregation methods, rather than inconsistencies in the results.

First, the difference between Table 2 (78.59) and Table 8 (80.12) reflects different checkpoint selection criteria. As described in Sec. 5.2.1, Table 2 evaluates cross-dataset generalization, so checkpoints are selected based on overall performance across multiple cross-domain validation sets. In contrast, Table 8 (Appendix B.3, Lines 1112–1114) evaluates in-domain localization, with checkpoints selected based on in-domain performance on the corresponding dataset. The optimal checkpoints for these two objectives do not necessarily correspond to the same training epoch. As training progresses, models may adapt more closely to the training distribution, improving in-domain performance without necessarily improving cross-domain generalization. This also explains the corresponding IoU differences for the other compared methods between Tables 2 and 8.

Second, regarding the difference between Table 2 (78.59) and Table 9 (76.96), both tables use the same model checkpoint but adopt different aggregation methods. As described in Appendix B.4 (Lines 1161–1163), Table 9 first calculates IoU separately for each of the 10 generator–operation subsets and then takes their unweighted average to examine localization difficulty across different generators and editing operations. In contrast, Table 2 reports the overall average IoU across the entire test set. Since the numbers of samples vary across subsets, these two aggregation methods naturally yield different results.

AuthorsOpenReview ↗

Response Part 3

W1 & Q3: Cross-Generator Generalization and Generator Overlap

Thank you for your valuable suggestions regarding the evaluation of generalization.

(1) Generator Overlap between Impostor and External Test Sets. We clarify that the four external test sets have limited generator overlap with Impostor, with PowerPaint being the only shared generator. BR-Gen uses LaMa, MAT, SDXL, BrushNet, and PowerPaint. SID-Set uses Latent Diffusion Model (LDM) for local manipulations. CocoGlide uses GLIDE, while AutoSplice uses DALL·E 2. In addition, these external test sets also differ in editing methods and data distributions. Together, these differences suggest that the cross-dataset evaluation in Table 2 provides meaningful evidence of the model’s ability to generalize across generators and data distributions.

(2) Performance Differences across External Test Sets. We agree with your observation that model performance varies substantially across different external test sets. Cross-dataset manipulation localization remains a major challenge in image manipulation detection and localization (IMDL), as different manipulation datasets often exhibit substantial differences in manipulation artifacts, editing methods, and data distributions. As shown in Table 2, other methods also exhibit substantial performance variations across target datasets. In particular, BR-Gen includes some background-replacement operations, whereas Impostor primarily focuses on local manipulations of foreground objects. This difference in manipulation types and spatial distributions may partly explain why all methods achieve relatively lower performance on BR-Gen. Despite this challenge, PANet achieves the highest IoU of 32.69% on BR-Gen and also ranks first on the other three external test sets. Thus, PANet’s strong cross-domain generalization relative to the compared methods is reflected not only in the aggregate results in Table 2, but more importantly in its consistently leading performance across all four external test sets.

(3) Leave-One-Generator-Out (LOGO) Evaluation. Finally, following your suggestion, we conducted additional LOGO experiments to more directly assess generalization to unseen generators while reducing potential confounding factors arising from cross-dataset differences in image sources, manipulation types, and annotation protocols. Specifically, in each run, we exclude all training samples from one generator, train the model using data from the remaining six generators, and evaluate it on the test subset corresponding to the excluded generator. Given the available time and computational resources, we selected the three baselines with the strongest cross-domain performance in Table 2 for comparison under the same LOGO protocol.

MethodFLUX.1 FillFLUX.1 KontextPowerPaintStep1X-EditQwen-Image-EditOmniEraserNano Banana ProAverage
NFA-ViT34.4160.4240.9169.6451.2919.5719.8842.30
RITA47.2360.7141.2564.2653.5923.7521.3944.60
Mesorch32.4152.5949.6671.6643.4818.4722.8841.59
PANet40.2366.6341.7377.9356.0124.1726.6247.62

Table B: Leave-One-Generator-Out (LOGO) localization performance on Impostor (IoU, %).

Table B shows that PANet achieves the best performance on five of the seven unseen generators, with an average IoU of 47.62% and absolute IoU gains of 3.02%, 5.32%, and 6.03% over RITA, NFA-ViT, and Mesorch, respectively. This suggests that PANet learns localization cues that generalize across generators, rather than relying solely on manipulation artifacts specific to the generators seen during training. As a complement to the cross-dataset evaluation in Table 2, the LOGO experiments provide more direct evidence of PANet's ability to generalize to unseen generators.

We have added an analysis of generator overlap between Impostor and the external test sets to Sec. 5.2.1 and included the LOGO experiments in Appendix B.2 to directly assess PANet’s generalization to unseen generators. All corresponding revisions are highlighted in red for ease of review.

AuthorsOpenReview ↗

Response Part 4

Q4: Reliability of Addition Masks

Thank you for your concern about the reliability of masks for the Add operation.

First, we clarify why we adopt post-generation segmentation for the Add operation. Adding a new object to an existing scene requires the object to blend naturally with its surroundings in terms of semantics, lighting, and physical relationships. Therefore, it is difficult to accurately define the final shape and boundaries of the added object using a predefined mask before generation. To address this, DisAgent first analyzes the scene and generates context-aware editing instructions. After image generation, we use Grounded SAM 2 to segment the added object and obtain its GT mask.

Second, although the Add operation uses post-generation segmentation, both the generated images and masks undergo DisAgent’s full quality validation pipeline, followed by the human verification stage. At this stage, two annotators independently compare the original image, manipulated image, and GT mask. They check for editing spillover beyond the intended region and assess whether the mask accurately delineates the added object. Samples with edits outside the intended region or segmentation errors are discarded, and only those approved by both annotators are retained.

Finally, to quantitatively assess the reliability of post-generation segmentation masks for the Add operation, we conducted an additional manual refinement study. We randomly selected 100 images from each of the four Add subsets, resulting in 400 samples. Annotators manually refined the masks for the added objects using the original images, manipulated images, and existing GT masks as references. The average IoU between the original GT masks and the refined masks was 94.1%, indicating high agreement between the two. Further inspection revealed that the discrepancies were mainly concentrated along the boundaries of SAM-segmented objects, reflecting minor boundary deviations rather than unintended edits beyond the intended region. These results further support the reliability of post-generation segmentation masks for the Add operation and suggest that automatic quality filtering and manual alignment review help control potential annotation errors.

Reviewer_TGAsOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Impostor, a benchmark for AI local edit detection and localization, containing 100,000 source-edited image pairs across seven generators, three operation types (addition, removal, and repainting), and multi-region edits. The data is built through DisAgent's perception, planning, generation, verification, and reflection pipeline; masks are obtained automatically and then manually checked for naturalness and alignment. The paper also proposes PANet, which combines residual and local phase cues with semantic features and predicts tampered regions via cross-attention, forensic–semantic consistency supervision, and regional contrastive learning. Experiments cover cross-dataset localization, image-level detection, module ablation, degradation robustness, and mechanism diagnosis.

Critical strengths

It provides a broad local-edit data resource that has gone through multi-stage quality checks. Impostor covers multiple recent editing models, different operations, and multi-region scenarios. Automatic quality filtering plus two-person review of naturalness and mask alignment support data quality; during splitting, the same source image and all its edited versions are placed in the same subset to reduce direct source-image leakage. Analyses of generators, edit area, and region count also make benchmark difficulty easier to interpret.

The model improvements are supported by fairly complete experiments and mechanistic evidence. PANet's main comparisons include recent dedicated localization methods, and report cross-dataset results, module ablations, and various post-processing degradations. The appendix checks the forensic–semantic consistency assumption via source images, shuffled masks, and low-frequency controls, and also compares local phase with other frequency-domain cues; this evidence helps understand the mechanism better than performance tables alone.

Critical weaknesses

The reliability of pixel-level masks still needs quantitative clarification. Masks come from predefined regions or post-edit object segmentation, and the paper provides automatic background consistency checks and manual alignment review, so it cannot be said that quality control is missing. However, object boundaries may not fully match the actual extent of editing effects, especially with shadows, edge blending, and small non-target changes. Existing semantic feature checks and manual pass rates are not enough to quantify such pixel-boundary errors, yet masks are the direct basis for localization training and evaluation.

Attribution of gains from training data sources is still insufficient. The appendix already compares multiple training sources, and results on some common targets support Impostor's advantage. However, when computing average scores, each source's own training dataset is excluded, so each source's average corresponds to a different test set; whether training sample size and budget are comparable is also not fully explained. Therefore, the existing average-score differences are not enough to separately distinguish the roles of data quality, diversity, and training scale, limiting quantitative interpretation of the data-construction contribution.

Localization false positives on real images are not yet sufficiently characterized. Localization IoU/F1 is mainly reported on already-tampered images; although image-level detection uses balanced real and tampered samples, those results cannot directly show how many erroneous regions would be produced on real images. For forensic applications, normal textures, edges, or imaging differences being misjudged as tampering also matters. The lack of such results limits judgment of the model's practical reliability.

Questions for authors

Do masks represent target objects or designated edit regions, or all pixels actually modified? How are shadows, boundary blending, and non-target detail changes handled? Can a small number of existing samples with manually refined masks, or boundary-tolerance evaluation, be used to quantify label error and its impact on localization results?

Can the existing training-source comparison table be re-aggregated directly on common target datasets not involved in training, and can training sample size and budget comparability be stated? If budgets or scales differ, please distinguish the effects of data quality and training scale; this question first needs clarification from existing results and does not require new large-scale training.

Can pixel false-positive ratio, erroneous-region statistics, or localization performance at a fixed image-level false-positive rate be added on the existing real-image test set? This would help determine whether the main localization gains come with more false positives on real content.

3 comments
AuthorsOpenReview ↗

Response Part 1

Dear Reviewer TGAs,

Thank you for your thoughtful and constructive feedback. We appreciate your recognition of Impostor’s value and PANet’s technical soundness. Below, we address your concerns and questions point by point.

W1 & Q1: Mask Definition and Quantitative Annotation Reliability

Thank you for your helpful suggestion.

First, we clarify that the GT masks in Impostor represent either newly added objects or designated editing regions, rather than pixel-wise differences between the original and manipulated images. Some generative editing models regenerate or reconstruct the entire image, which may introduce slight pixel-level changes in non-target regions even when their visual content remains unchanged. Conversely, some pixels within the target region may remain unchanged or change only slightly. GT masks derived from pixel differences may therefore include background noise while missing parts of the target region. In contrast, masks defined by the target object or designated editing region provide a more stable representation of the intended manipulation area.

Second, regarding shadows, boundary blending, and subtle changes in non-target regions, we distinguish minor reconstruction differences that do not materially affect visual content from actual edits extending beyond the target region. To address these issues, we combine feature-based constraints, LVLM-based realism assessment, and manual alignment review during dataset construction. Specifically, MA-BC (Appendix E.5) constrains the editing scope based on feature consistency rather than strict pixel-wise equality. It tolerates minor reconstruction differences that do not affect background content while filtering non-target modifications that cause feature inconsistencies. In addition, LVLM-based realism assessment checks for visible blending artifacts and inconsistencies in lighting and textures around the boundaries of edited regions. Finally, two annotators independently examine the original image, edited image, and GT mask. In addition to assessing image naturalness, they verify the spatial alignment between the actual editing region and the annotated region, and reject samples with visible shadow spillover, boundary misalignment, or non-target content modifications. Together, these checks reduce the risk of discrepancies between the actual editing region and the GT mask.

To further quantify the accuracy of GT mask annotations and assess the impact of potential annotation errors on localization performance, we randomly sampled 100 images from each of the 10 generator–operation subsets in the Impostor test set, resulting in 1,000 samples. Annotators manually refined the masks by comparing the original images, manipulated images, and existing GT masks. We then calculated the IoU between the original and manually refined GT masks to assess annotation accuracy. Furthermore, keeping model predictions fixed, we calculated localization IoU separately against the original and refined GT masks on the same 1,000 images.

MetricMVSSNetTruForMesorchM2SFormerNFA-ViTRITAPANet
Original GT IoU ↑\uparrow52.6448.1671.7457.1370.1968.1478.76
Refined GT IoU ↑\uparrow53.19 (+0.55)48.14 (−0.02)72.40 (+0.66)57.14 (+0.01)70.87 (+0.68)68.09 (−0.05)79.47 (+0.71)

Table A. Pixel-level localization performance using original and manually refined annotations (IoU, %). Parentheses indicate changes in IoU relative to the original annotations.

The results show that the mean IoU between the original GT masks and the manually refined masks is 96.51%, indicating high agreement between the original annotations and the refined references. Further inspection reveals that most annotation discrepancies are concentrated near the boundaries of SAM-segmented target objects, with only minor discrepancies in non-target regions. Furthermore, re-evaluating all methods using the refined masks yields absolute changes in localization IoU of no more than 0.71%, indicating that annotation refinement has a limited impact on localization performance.

These results support the reliability of Impostor's pixel-level annotations. We have added this quantitative analysis to Appendix C.3 of the revised manuscript, with the new content highlighted in red.

AuthorsOpenReview ↗

Response Part 2

W2 & Q2: Fair Comparison of Training Sources on Common Test Sets.

We thank the reviewer for this suggestion.

First, regarding the fairness of training data scale and computational budgets, the experiments in Table 7 used the same number of training samples and identical training configurations across all training sources. Specifically, we randomly sampled 88,000 manipulated images from each of the SID-Set and BR-Gen training sets to match the number of manipulated images in the Impostor training set. The number of authentic images was also kept constant across all training sources. Furthermore, for each localization model, we kept the network architecture, optimization settings, batch size, and number of training epochs identical across all training sources, ensuring a fair comparison.

Second, following the reviewer’s suggestion, we re-aggregated the existing results in Table 7 (Appendix B.1) using CocoGlide and AutoSplice as common unseen test sets for all three training sources. We compute the arithmetic means of IoU and F1 across these two datasets, with equal weight for each dataset. The results are reported below.

MethodTraining SourceMean IoU ↑\uparrowMean F1 ↑\uparrow
M2SFormerImpostor53.8263.55
M2SFormerBR-Gen51.9359.47
M2SFormerSID-Set41.9648.22
MesorchImpostor67.2774.60
MesorchBR-Gen63.0869.03
MesorchSID-Set38.6943.37
PANetImpostor70.5977.59
PANetBR-Gen67.0572.98
PANetSID-Set49.2052.95

Table B: Cross-dataset generalization on common unseen test sets (CocoGlide and AutoSplice).

The results show that all three localization models still achieve the best cross-domain generalization when trained on Impostor, under matched training sample sizes, training configurations, and common unseen test sets.

We have further clarified the training settings of Table 7 and re-aggregated the results in Appendix B.1 of the revised manuscript to ensure a more direct and fair comparison across training sources.

AuthorsOpenReview ↗

Response Part 3

W3 & Q3: False-Positive Localization on Authentic Images

Thank you for your valuable feedback. We agree that false positives on authentic images are an important factor in assessing the practical reliability of IMDL models. Therefore, we provide pixel-level false-positive analysis on authentic images across multiple test sets and further evaluate localization performance at a fixed image-level false-positive rate.

Specifically, we report two metrics: (1) Pixel FPR, the proportion of pixels in authentic images incorrectly predicted as manipulated, computed directly from the original segmentation predictions; and (2) IoU @ 4% Image FPR, manipulation localization performance evaluated at a fixed image-level FPR of 4%.

For this analysis, we include only methods that provide both image-level detection scores and pixel-level localization outputs. All methods use the same model checkpoints as in Table 2, without additional training. For IoU @ 4% Image FPR, we calibrate each model's image-level decision threshold on an independent authentic-image validation set for each dataset, targeting an image-level FPR of approximately 4%. We then jointly evaluate detection and localization on the corresponding manipulated test set. Images classified as authentic are assigned empty masks, while the original segmentation predictions are retained for images classified as manipulated. Mean IoU is computed over all manipulated test images.

DatasetImpostorImpostorImpostorSID-SetSID-SetSID-SetBR-GenBR-GenBR-Gen
MethodPixel FPR (%) ↓\downarrowOriginal IoU ↑\uparrowIoU @ 4% Image FPR (Val.-Calibrated) ↑\uparrowPixel FPR (%) ↓\downarrowOriginal IoU ↑\uparrowIoU @ 4% Image FPR (Val.-Calibrated) ↑\uparrowPixel FPR (%) ↓\downarrowOriginal IoU ↑\uparrowIoU @ 4% Image FPR (Val.-Calibrated) ↑\uparrow
NFA-ViT0.310467.5166.12 (−1.39)14.2231.5021.12 (−10.38)13.2724.4121.31 (−3.10)
MVSSNet0.654850.0041.18 (−8.82)26.2227.8919.91 (−7.98)37.1312.979.49 (−3.48)
TruFor1.317151.1740.61 (−10.56)24.1822.3113.36 (−8.95)25.1916.8511.91 (−4.94)
ObjectFormer1.851554.1951.61 (−2.58)16.7724.8210.62 (−14.20)14.1330.3022.65 (−7.65)
PANet0.101878.5977.63 (−0.96)9.1246.4340.56 (−5.87)8.2732.6930.53 (−2.16)

Table C: False-positive analysis on authentic images and localization performance at 4% image-level FPR. Parentheses indicate absolute IoU drops (%) relative to the original results.

As shown in Table C, PANet achieves the lowest Pixel FPR among all compared methods on Impostor, SID-Set, and BR-Gen. Moreover, under the approximately 4% image-level FPR constraint, PANet achieves the best localization performance on all three datasets, with the smallest IoU drop in each case.

Thus, PANet’s localization advantage does not come at the cost of increased false positives on authentic images.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response: Thanks, Recognized Contributions, Key Clarifications, and Additional Experiments

We sincerely appreciate all reviewers’ time and efforts in reviewing our paper. We are glad that the reviewers recognize the benchmark value of Impostor, the closed-loop data construction framework of DisAgent, and the methodological design and empirical performance of PANet. In response to the reviewers’ comments, we further clarify the experimental evidence in the original paper and provide several targeted quantitative analyses. Below, we summarize the contributions recognized by reviewers, the key clarifications, and the additional experiments and analyses in this rebuttal.

Contributions Recognized by Reviewers

The main contributions recognized by the reviewers are summarized as follows:

  • Benchmark Value. Reviewers recognize Impostor as a valuable and challenging benchmark for AIGC-based image manipulation detection and localization. [TGAs, cuLG, gof6]
  • Agent-Based Construction. Reviewers recognize DisAgent's closed-loop data construction framework, which combines autonomous planning, quality validation, and iterative feedback, going beyond simple sequential generation. [gof6]
  • Methodology and Evaluation. Reviewers recognize the technical soundness of PANet and acknowledge that its effectiveness is supported by comprehensive experiments and mechanistic evidence. [TGAs, gof6]
  • Transferability as a Training Source. The reviewer recognizes the cross-domain generalization value of Impostor as a training source, noting its consistent advantages across multiple localization architectures. [cuLG]

New Clarifications, and New Experiments

We also thank all reviewers for their insightful and constructive suggestions, which have helped us further improve our paper. In addition to the point-by-point responses below, we summarize the key clarifications and additional experiments included in this rebuttal in response to the reviewers’ suggestions.

New Clarifications

  • Mask Definition and Annotation Reliability. Clarification of the GT Mask definition and the reliability of the mask annotation process. [TGAs, cuLG]
  • Experimental Protocols and Fair Comparisons. Clarification of the test-time input setting, evaluation protocols, and the fairness of comparisons across different training sources. [TGAs, cuLG]
  • Cross-Domain and Cross-Generator Generalization. Clarification of the generators used in external benchmarks, their overlap with Impostor, and performance differences across test domains. [cuLG]
  • PANet and DisAgent Contributions. Clarification of PANet’s core methodological insight and the independent contributions of DisAgent’s closed-loop mechanisms. [cuLG, gof6]
  • Visual Realism and Shortcut Learning. Clarification of the visual realism assessment of Impostor and its relationship to shortcut learning. [gof6]
  • Framework Extensibility and Responsible Dataset Release. Clarification of DisAgent’s extensibility to emerging image editing models and the potential risks of dataset release, along with mitigation measures. [gof6]

Additional Experiments and Analyses

  • Mask Annotation Reliability. Evaluation of the agreement between original and manually refined masks, as well as the impact of mask annotation errors on localization performance. [TGAs, cuLG]
  • Localization Reliability. Analysis of pixel-level false positive rates (Pixel FPR) on authentic images and evaluation of localization performance at a fixed image-level FPR calibrated on an authentic-image set. [TGAs]
  • Difference Baselines and Generator Generalization. Localization baselines based on pixel-level and frozen-feature differences using paired source images, together with a leave-one-generator-out (LOGO) evaluation across seven generators. [cuLG]
  • Visual Realism and Shortcut Learning. LVLM-based recognition evaluation on authentic images, and controlled training comparisons between High- and Low-Realism images. [gof6]
  • Generalization to Human-Edited Manipulations. Zero-shot localization evaluation on traditional image manipulation benchmarks to assess PANet’s generalization ability. [gof6]
  • FSC Contribution Visualization. Visualization of results with and without FSC to illustrate its role in localizing manipulated regions. [gof6]
  • DisAgent Ablation. Ablation studies of DisAgent across different generators. [gof6]

We have further revised the manuscript to clarify the methodology, experimental protocols, and analyses of annotation reliability and generalization. We have also expanded the Ethics Statement to address the responsible release of the dataset. All revisions are highlighted in red for ease of review. We sincerely thank the reviewers again for their valuable suggestions.

Dear reviewers,

As the discussion period approached the end, please take a moment to read the authors’ responses and share your thoughts on any remaining points of discussion. If the discussion has changed your assessment, please update your review accordingly.

Thank you for your time and contributions!

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.