- 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
- 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.
- 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.
- 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.
- 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
- 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.
- 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.
- 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.
- 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
- How much of the improvement in PANet comes from FSC learning compared with the local forensic feature extraction module?
- How well does PANet generalize to real-world manipulated images created by human editors rather than AIGC models?
- Could DisAgent be applied to continuously expand the benchmark as new image editing models emerge?
- What mechanisms will be used to mitigate potential misuse of the released manipulation dataset?
4 comments
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).
| LVLM | Real Image Recognition Accuracy (%) |
|---|---|
| Qwen3-VL-235B-A22B-Instruct | 92.5% |
| FakeShield | 94.1% |
| Gemini 3 Pro | 93.2% |
| GPT-5.4 | 99.5% |
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.
| Method | Training Subset | AutoSplice | BR-Gen | CocoGlide |
|---|---|---|---|---|
| TruFor | Low-Realism | 58.17 | 58.25 | 52.13 |
| TruFor | High-Realism | 67.16 | 67.00 | 54.85 |
| NFA-ViT | Low-Realism | 70.41 | 59.15 | 54.13 |
| NFA-ViT | High-Realism | 78.94 | 69.40 | 59.24 |
| PANet | Low-Realism | 69.21 | 64.46 | 57.23 |
| PANet | High-Realism | 81.82 | 73.22 | 63.65 |
Table C: Effect of training data visual realism on cross-domain image-level detection (Balanced Accuracy, % ).
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.
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.
| Method | CASIA V1(918) | CASIA V2(3,000) | Columbia(180) | Coverage(91) | IMD2020(2020) | Average |
|---|---|---|---|---|---|---|
| MVSSNet | 14.64 | 9.62 | 18.18 | 9.47 | 7.46 | 11.87 |
| TruFor | 15.01 | 12.46 | 11.79 | 12.43 | 12.37 | 12.81 |
| Mesorch | 29.10 | 25.32 | 19.26 | 9.97 | 18.05 | 20.34 |
| M2SFormer | 30.08 | 20.40 | 22.31 | 12.73 | 19.36 | 20.98 |
| NFA-ViT | 23.00 | 23.13 | 21.98 | 10.44 | 16.92 | 19.09 |
| PSCC-Net | 14.99 | 15.50 | 12.92 | 11.49 | 17.75 | 14.53 |
| RITA | 29.84 | 26.07 | 25.50 | 12.39 | 21.76 | 23.11 |
| PANet | 31.93 | 28.14 | 18.72 | 15.60 | 23.63 | 23.60 |
Table D: Zero-shot generalization from Impostor to conventional image manipulation benchmarks (IoU, % ).
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.
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.
| Setting | Validation | Reflection | Image Det. (%) | Pixel IoU (%) |
|---|---|---|---|---|
| Sequential Pipeline | ✗ | ✗ | 33.67 | 14.56 |
| w/o Validation | ✗ | ✓ | 29.63 | 13.76 |
| w/o Reflection | ✓ | ✗ | 21.70 | 12.70 |
| Full DisAgent | ✓ | ✓ | 19.37 | 12.28 |
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.
| Generator | Sequential | Validation Only | Reflection Only | Full DisAgent |
|---|---|---|---|---|
| Nano Banana Pro | 23.63 | 19.11 | 22.76 | 14.08 |
| Step1X-Edit | 33.75 | 23.33 | 31.40 | 20.81 |
| FLUX.1 Kontext | 39.74 | 27.04 | 34.84 | 24.62 |
| Qwen-Image-Edit | 45.93 | 30.78 | 39.83 | 25.87 |
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.
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.
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.