acceptodds
Desk rejected at ICLR 2027

TRACER: Dual Evidence Modeling for Generalizable Image Forgery Detection

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent advances in image generation and editing have led to increasingly diverse and realistic image forgeries, making reliable detection more challenging. However, existing forensic methods are typically designed for specific forgery forms and rely on type-specific artifacts, such as boundary and noise inconsistencies in locally manipulated images or structural and frequency irregularities in AI-generated images, making them difficult to extend to other forgery scenarios. To address this limitation, we propose TRACER, a generalizable image forgery detection framework for diverse forgery forms. Our key insight is that forgery processes introduce anomalous deviations from real-image distributions, which can be primarily characterized from two views, fine-grained and logical. Accordingly, TRACER adopts a dual-stream architecture that models each evidence view with a dedicated stream. To promote their collaboration, Bidirectional Evidence Sharing facilitates information exchange between streams, while Hybrid Layer Interaction controls where such exchange occurs across feature depths. Progressive Evidence Training then gradually activates cross-stream interaction. Across benchmarks covering several forgery scenarios, TRACER achieves state-of-the-art performance, improving average evidence localization mIoU by 8.3% and judgment accuracy by 22.5%. It also reaches 92.6% zero-shot accuracy on synthetic images and videos from recent generators. These results show the potential of dual evidence modeling for accurate and generalizable image forgery detection.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Forgery Detection, Vision-Language Models

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

This work targets the unified image forgery evidence localization and analysis task, aiming to detect three forgery categories: traditional local manipulation, generative editing, and fully AI-synthesized images. The proposed TRACER framework consists of two collaborative evidence streams: a fine-grained stream built on SAM for local visual anomaly localization, and a logical semantic stream based on Qwen3-VL-4B for reasoning scene plausibility and generating natural-language explanations. A Bidirectional Evidence Sharing (BES) cross-attention module enables mid-layer feature exchange between two branches, combined with hybrid layer interaction and progressive evidence training to stabilize the joint optimization. The model outputs both pixel-wise anomaly masks and textual authenticity judgments with anomaly analysis. Extensive experiments across multiple forgery benchmarks and unseen generative models demonstrate promising overall performance, yet several critical concerns remain regarding task definition, empirical validation and explanation fidelity.

Critical strengths
  • The research problem is practically valuable. This work unifies detection for traditional tampering, generative editing and fully AI-generated images under one framework, instead of handling each forgery type separately.

  • The dual-evidence modeling paradigm is intuitive and well-motivated. It complements low-level forensic traces and high-level semantic logical constraints, with bidirectional cross-branch information exchange to boost localization and classification.

  • The authors carefully design interaction strategies and progressive training scheduling to mitigate distribution mismatch between two pre-trained backbones, not only proposing an architecture but also investigating training stability.

  • The evaluation scope is comprehensive, covering localization, binary forgery classification, natural language explanation, and zero-shot tests on unseen image/video generators.

  • Joint mask prediction and textual interpretation provides human-readable evidence, which is more suitable for forensic assistant scenarios than standalone classification scores.

Critical weaknesses
  • Limited methodological novelty: The core BES module is essentially cross-attention KV-sharing across different feature spaces. The unique contribution of the proposed dual-evidence mechanism has not been fully disentangled from stronger backbones, enlarged training data and complex training pipelines. Simple fusion baselines and parameter-matched ablations are missing.

  • Inconsistent supervision targets for anomaly masks: The mask annotation definition differs drastically across forgery types. Masks denote edited regions for traditional tampering/generative editing, but only manually marked semantic anomaly regions for fully AI-generated images. The unified mask supervision may learn dataset-specific annotation biases rather than a universal anomaly concept. Inter-annotator agreement of anomaly labels is not validated.

  • Insufficient robustness and real-world generalization: Strong performance on unseen generators does not guarantee robustness against post-processing such as compression, blur, filters and multi-stage mixed editing. The model degrades significantly under heavy image degradation. Video evaluation only uses static frames and cannot verify temporal forgery detection.

  • Unreliable textual explanation evaluation: The EQS metric relies on a single LLM judge. High EQS scores cannot prove the explanations are factually consistent or causally faithful to visual evidence; hallucinated false evidence claims are not quantified. Human validation and cross-judge verification are absent.

  • Uneven performance across sub-datasets: Aggregated average metrics mask large performance gaps. Especially for fully AI-generated benchmarks, the localization F1 score remains very low. Detailed per-class precision, recall and error type analysis are inadequate.

Questions for authors

Please refer to the weaknesses.

6 comments
AuthorsOpenReview ↗

Response to Reviewer cxoP (Part 1)

We sincerely thank you for your careful reading and valuable suggestions. Below are our responses to your questions and concerns.

Response to W1:

We appreciate your interest in BES, particularly your suggestions for additional fusion baselines and parameter-scale ablations. We address your concerns by explaining the interaction mechanism and presenting additional experiments.

Our BES uses a gate that gradually opens during training to enable progressive bidirectional evidence sharing. In common interaction methods such as ViLBERT [1] and CALM [2], cross-attention uses one network’s Q to query another’s K/V. Our design has three main differences. First, a gate on the cross-stream interaction path initially suppresses cross-stream contributions, allowing each stream to rely mainly on its own features. The gate gradually opens during training until fully enabled, helping limit early cross-stream interference. Second, we introduce learnable bridging projectors in both directions. These projectors learn to map the other stream's K/V into the current stream's feature space, helping the two experts gradually adapt to each other's different feature representations. Third, each stream attends jointly to its own K/V and the projected cross-stream K/V, allocating attention between them based on its current Q. This allows the logical stream to use texture and boundary cues and the fine-grained stream to use object relations and commonsense information, improving their respective evidence modeling. The ablation results in Table 4 of the main paper further support these benefits in both directions.

Regarding the effects of training data and the training procedure on performance, all baselines in Table 1 of the main paper use the same training data as TRACER. TRACER still achieves better overall performance. Some results are shown below, and the complete results are available in Table 1 of the main paper:

MethodCASIAv1+CocoGlideSynthScars
mIoU / F1mIoU / F1mIoU / F1
TruFor.753 / .685.512 / .475.538 / .314
CAT-Net v2.753 / .718.610 / .502.518 / .210
SIDA.693 / .511.554 / .503.534 / .278
TRACER.919 / .893.655 / .636.560 / .329

Table 7 in the main paper further shows that TRACER trained directly, even without progressive gating or the subsequent capability refinement stage, already outperforms SIDA overall, while the full training strategy further improves on these results. The results are shown below:

MethodStage 1GateStage 2CASIAv1+SynthScars
mIoU / AccmIoU / Acc
SIDA–––.693 / .316.534 / .182
TRACER✓✗✗.903 / .561.534 / .938
TRACER✓✓✓.919 / .946.560 / .999
AuthorsOpenReview ↗

Response to Reviewer cxoP (Part 2)

To further clarify the respective effects of interaction design and parameter scale, we added controlled experiments. First, we compare the following three fusion baselines while keeping the dual-stream backbones, bridging-layer parameter counts, training data, and training procedure the same:

  1. Final-layer Feature Fusion concatenates each stream's final-layer features with MLP-projected features from the other stream.
  2. Layer-wise Feature Fusion uses the same interaction as the first baseline, but applies Hybrid Layer Interaction to exchange information at multiple depths.
  3. Layer-wise Cross-Attn uses each stream's own Q to attend only to the other stream's K/V at the selected interaction layers.

We also replace only the logical expert Qwen3-VL-4B with Qwen3-VL-2B for the parameter-scale ablation, keeping the architecture and all mechanisms unchanged. Results are shown below:

SettingCASIAv1+CocoGlideSynthScarsSITD
mIoU / AccmIoU / AccmIoU / AccAcc
Final-layer Feature Fusion (4B).848 / .866.605 / .691.476 / .898.948
Layer-wise Feature Fusion (4B).880 / .891.605 / .727.550 / .954‾\underline{.954}.950
Layer-wise Cross-Attn (4B).590 / .460.485 / .428.471 / .260.922
Layer-wise BES (2B).906‾\underline{.906} / .916‾\underline{.916}.643‾\underline{.643} / .741‾\underline{.741}.557‾\underline{.557} / .999.962‾\underline{.962}
Layer-wise BES (4B).919 / .946.655 / .745.560 / .999.989

The results show that Layer-wise BES achieves the best overall performance among the fusion baselines. Layer-wise Cross-Attn performs markedly worse, possibly because attending only to cross-stream K/V disrupts each expert’s evidence analysis. BES instead jointly attends to its own and projected cross-stream K/V, enabling more effective use of both evidence types. Layer-wise Feature Fusion’s overall advantage over Final-layer Feature Fusion also supports the benefits of interaction at multiple depths. The 2B TRACER variant further outperforms the 4B fusion baselines, suggesting that its gains are not solely due to backbone size. These results support the effectiveness of dual-evidence interaction. We have added Table 5 and the corresponding analysis under "Different Fusion Strategies" in Section 4.2 of the revised manuscript.

[1] J. Lu, D. Batra, D. Parikh, et al. ViLBERT: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Advances in Neural Information Processing Systems, vol. 32, pp. 13–23, 2019.

[2] R. Bansal, B. Samanta, S. Dalmia, et al. LLM augmented LLMs: Expanding capabilities through composition. In International Conference on Learning Representations, 2024.

AuthorsOpenReview ↗

Response to Reviewer cxoP (Part 3)

Response to W2:

Thank you for highlighting that masks for different forgery types differ in their annotation criteria and spatial granularity.

TRACER's unified design mainly lies in jointly modeling fine-grained and logical evidence to identify anomalies that deviate from the real-image distribution across different forgery types. Masks for different forgery types all provide supervision for this goal. Although their forms are not identical, they all indicate anomalous regions in forged images that deviate from the real-image distribution, providing spatial supervision for localizing visible anomaly cues related to forgery. Our framework aims to use this supervision to learn anomaly cues that differ from real images across different forgery processes. Predicted masks and textual analysis jointly present the evidence supporting authenticity assessment, making detection results easier to interpret.

However, we also recognize that different annotation methods may still differ in the exact regions covered by their supervision. This is also an important reason why TRACER combines masks with textual explanations, which can provide additional context when anomaly cues lack clear spatial boundaries, as illustrated in Appendix Figure 11. In future work, we will further explore more consistent approaches to anomaly supervision and evaluation.

Response to W3:

Thank you for your interest in the model's robustness. We would like to clarify that Table 10 in the appendix already evaluates robustness to several common image post-processing operations and perturbations at different strengths. Building on this evaluation, we added robustness experiments on ArtiBench [3], which includes generators unseen during training. The table below summarizes the results on four benchmarks:

BenchmarkMetricNo PerturbationJPEG Compression QF=90JPEG Compression QF=80JPEG Compression QF=70Gaussian Noise σ=0.02Gaussian Noise σ=0.06Gaussian Noise σ=0.10Gaussian Blur Ksize=3Gaussian Blur Ksize=5Gaussian Blur Ksize=7Gamma Correction γ=0.7Gamma Correction γ=1.3
CASIAv1+mIoU.919.903.876.850.899.833.779.819.772.717.909.915
CASIAv1+Acc.946.925.915.907.922.783.760.711.673.770.913.944
SynthScarsmIoU.560.543.545.544.547.534.517.557.560.559.551.550
SynthScarsAcc.9991.000.999.996.962.800.7221.0001.0001.000.988.999
ArtiBenchmIoU.499.486.485.485.478.473.471.493.498.501.480.483
ArtiBenchAcc.948.960.972.962.808.526.408.990.994.996.922.960
COCOAcc.874.854.836.790.823.935.879.504.317.207.871.816

The results show that TRACER is generally stable under JPEG compression and gamma correction, while stronger Gaussian noise and blur can lead to performance decreases on some benchmarks. Strong blur may remove natural high-frequency details, making real images resemble generated ones. Overall, it shows good robustness.

In addition, we added Figure 8 to the appendix of the revised manuscript to compare the robustness of different models. All comparison methods use the same training data as TRACER and are evaluated under the same perturbation conditions. On CASIAv1+ and SynthScars, TRACER generally outperforms the comparison methods in mIoU and Acc across perturbation strengths, further supporting its robustness under the tested conditions.

Regarding the use of only static frames for video evaluation, TRACER currently focuses on image forgery detection. The existing video experiments therefore extract static frames and detect forgery in each frame independently, without modeling temporal relationships between frames. In future work, we will explore temporal information to extend the framework to video forgery detection.

We have updated Table 10 under "Robustness Experiments" in Appendix A.2 of the revised manuscript to include the robustness results on ArtiBench. We have also added Figure 8 and the corresponding analysis to compare robustness across models.

[3] J. Park, M. Ahn, M. Kim, et al. See and fix the flaws: Enabling VLMs and diffusion models to comprehend visual artifacts via agentic data synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 35810–35820, 2026.

Reviewer_UBQXOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The manuscript proposes TRACER, a generalizable image forgery detection and localization framework designed to handle unseen forgery types and datasets. The authors claimed that forgery processes introduce anomalous deviations from real-image distributions, which can be categorized into fine-grained and logical artifacts. TRACER thereby adopts a dual-stream architecture to model these two types of forensic clues. In addition, the authors introduce Bidirectional Evidence Sharing, Hybrid Layer Interaction, and Progressive Evidence Training to further facilitate the fusion of the two feature representations. Experimental results demonstrate that the proposed method achieves promising generalizability.

Critical strengths

This manuscript provides extensive generalizability evaluations and ablation studies. The generalizability experiments demonstrate promising performance on unseen datasets and manipulation types. Compared with previous LLM-based image forgery localization methods, the proposed method uses only 4B parameters, demonstrating its parameter efficiency. Furthermore, the ablation studies validate the effectiveness of the designed components.

Critical weaknesses

The authors claim that forgery processes introduce anomalous deviations from real-image distributions, which can be categorized into fine-grained and logical artifacts. However, no statistical evidence is provided to support this claim. Further analysis of the distributions of real and forged images is necessary to substantiate this argument. Moreover, the predefined categorization into two types of artifacts is not sufficiently convincing, as the authors provide little justification for this distinction. This concern is particularly important because the proposed categorization serves as the primary motivation for the entire framework.

The novelty of the proposed framework is limited. Both the dual-stream architecture and progressive training strategy have been extensively explored in previous image forgery localization studies. Similarly, the proposed cross-attention mechanism for feature fusion is not particularly novel in this field. Overall, the manuscript lacks sufficient methodological novelty and a well-supported motivation. Therefore, I do not believe that the manuscript meets the acceptance standards of ICLR.

Questions for authors

Please see the critical weaknesses.

7 comments
AuthorsOpenReview ↗

Response to Reviewer UBQX (Part 1)

We sincerely thank you for your careful review and constructive feedback. We address your questions and concerns below.

Response to W1:

Thank you for your comments on our core motivation. We would like to clarify that the fine-grained and logical views do not constitute an exhaustive or mutually exclusive classification of forgery anomalies. They are two key perspectives we chose for analyzing anomaly evidence, focusing on visual details and the plausibility of scene content, respectively. A forged image need not exhibit both types, and other forms of anomalies may also exist. Our focus is whether these views help distinguish real from forged images and whether joint modeling improves detection. We have clarified the wording and added statistical analyses and visualizations in the revised manuscript to address these questions.

Our motivation stems from the different constraints on image formation and scene content: textures, boundaries, and noise are closely tied to the imaging process, while object structures, spatial relations, and lighting consistency are governed by scene semantics and physical laws. Representative detection methods also reflect these perspectives. For example, TruFor [1] captures subtle imaging traces through noise-sensitive fingerprints, and FakeShield [2] examines perspective relationships and lighting and shadow consistency. Despite their different designs, the cues used by these methods can be understood through fine-grained and logical views.

To illustrate the role of fine-grained and logical evidence in distinguishing real from forged images and the gains from combining them, we statistically analyzed forged images and their corresponding originals across three benchmarks covering different forgery types. We constructed the fine-grained indicator set F from eight features covering residual statistics, local residual co-occurrence deviations, and SIFT-based repetition cues. We derived the logical indicator set L from VLM log-odds scores on questions about object structure, physical relations, object interactions, and optical consistency. Under a common evaluation protocol, we trained separate logistic regression classifiers using F, L, and F+L to compare their ability to distinguish real from forged images. Experimental details are provided under “Statistical Analysis of Dual Anomaly” in Appendix A.2.

The newly added Figure 9 in Appendix A.2 summarizes our statistical results. Figure 9(a) compares the standardized distributions of two representative anomaly indicators on CocoGlide. Forged images have higher mean scores on both indicators, although the distributions overlap, suggesting that both cues are informative but neither alone provides perfect separation. Figure 9(b) compares the ROC curves of F, L, and F+L across decision thresholds, while Figure 9(c) shows that joint use improves AUROC across all three benchmarks. On CocoGlide, AUROC increases from 0.681 for the stronger individual setting to 0.700. These results show that both fine-grained and logical views capture differences between real and forged images, and that jointly using these views improves detection performance. The corresponding AUROC values are summarized in the table below:

DatasetFine-grained Only (F)Logical Only (L)Joint (F+L)
CocoGlide.576 (-.124).681 (-.019).700
COVERAGE.705 (-.038).525 (-.218).743
B-Free.816 (-.049).603 (-.262).865

To further verify whether TRACER's two streams focus on fine-grained and logical evidence, respectively, we added Figure 5 and the computation details to Section 4.3 of the revised manuscript. For the same forged image, we use gradient backpropagation from the final localization score to visualize both streams’ intermediate-layer spatial responses. As shown in the figure, the two streams have different areas of focus at intermediate layers: the fine-grained stream focuses more on local details such as the deer's outline and antlers, while the logical stream focuses more on object regions such as the deer and the person. These differences in spatial focus are consistent with the intended roles of the two streams in emphasizing local appearance and object context, respectively.

These results provide empirical support for jointly modeling fine-grained and logical evidence. We have further clarified this point in the revised Abstract and Introduction and hope this addresses your concern.

[1] F. Guillaro, D. Cozzolino, A. Sud, et al. TruFor: Leveraging all-round clues for trustworthy image forgery detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 20606–20615, 2023.

[2] Z. Xu, X. Zhang, R. Li, et al. FakeShield: Explainable image forgery detection and localization via multi-modal large language models. In International Conference on Learning Representations, 2025.

AuthorsOpenReview ↗

Response to Reviewer UBQX (Part 2)

Response to W2:

We would like to clarify that TRACER’s central contribution is a unified forgery detection framework covering multiple forms of image forgery. While existing methods often focus on specific forgery types or evidence sources, TRACER jointly models fine-grained and logical evidence using two experts with different pretraining objectives. The core goal of its dual-stream architecture, BES, and progressive training is to enable evidence sharing and collaboration between the two experts. Below, we explain these contributions through the design details and controlled experiments.

First, our dual-stream design emphasizes collaboration between experts during feature extraction. Existing work focuses more on integrating representations through adapters after the encoders have extracted features. For example, TranX-Adapter [3] uses an adapter to fuse expert features into visual tokens for a language model, while ForgerySleuth [4] combines image features and language model outputs in a visual decoder for localization. These methods primarily integrate expert outputs for downstream analysis. In contrast, TRACER instead retains each expert’s backbone and output path while connecting selected intermediate layers, aiming to preserve their capabilities and let each use the other’s evidence during anomaly discovery and analysis.

Second, BES enables progressive bidirectional information sharing between the two experts. ViLBERT [5] and CALM [6] use one network’s Q to query another’s K/V. BES combines three designs. First, a gate gradually opens to full strength during training, initially favoring each stream’s own features to limit early cross-stream interference. Second, bidirectional learnable projectors map cross-stream K/V into the receiving stream’s feature space, helping the experts adapt to each other’s representations. Third, each stream retains its native Q/K/V and jointly attends to its own and projected cross-stream K/V based on its current Q. The updated representations pass to subsequent layers, allowing the logical stream to use texture and boundary cues and the fine-grained stream to use object and scene context. The ablations in Table 4 of the main paper support these bidirectional benefits.

To evaluate BES against other fusion methods, we compared it with Feature Fusion and Cross-Attn at the same interaction layers. BES achieves better overall performance, suggesting that it is better suited to our task. The results are shown below, with detailed analysis provided under “Different Fusion Strategies” in Section 4.2 of the revised manuscript.

SettingCASIAv1+CocoGlideSynthScarsSITD
mIoU / AccmIoU / AccmIoU / AccAcc
Feature Fusion.880 / .891.605 / .727.550 / .954.950
Cross-Attn.590 / .460.485 / .428.471 / .260.922
BES.919 / .946.655 / .745.560 / .999.989

Third, progressive training helps experts with different pretraining objectives and feature distributions adapt to each other. Full-strength interaction from the outset may introduce cross-stream interference before this adaptation occurs. We therefore initially limit cross-stream contributions through gating and gradually strengthen interaction during joint optimization. Appendix Figure 7 provides further analysis of training stability and convergence. Related strategies mainly schedule learning tasks, trainable modules, or training data through staged visual and language training or data augmentation curricula. Our gating instead aims to promote gradual adaptation and collaboration between experts.

Together, TRACER’s designs enable collaborative dual-evidence detection across diverse forgery scenarios. Other reviewers have also recognized these designs: Reviewer cxoP also recognized the dual-evidence motivation and the role of progressive training in cross-stream adaptation and stability, and Reviewer jbQv acknowledged the benefits of bidirectional sharing for assessment and localization. We hope these explanations address your concerns.

[3] W. Wang, Y. Huang, J. Xu, et al. TranX-Adapter: Bridging artifacts and semantics within MLLMs for robust AI-generated image detection. In Proceedings of the 43rd International Conference on Machine Learning, vol. 306, pp. 127025–127039, 2026.

[4] Z. Sun, H. Jiang, H. Chen, et al. ForgerySleuth: Empowering multimodal large language models for image manipulation detection. In Advances in Neural Information Processing Systems, vol. 38, 2025.

[5] J. Lu, D. Batra, D. Parikh, et al. ViLBERT: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Advances in Neural Information Processing Systems, vol. 32, pp. 13–23, 2019.

[6] R. Bansal, B. Samanta, S. Dalmia, et al. LLM augmented LLMs: Expanding capabilities through composition. In International Conference on Learning Representations, 2024.

AuthorsOpenReview ↗

Official Comment by Authors

We sincerely appreciate the reviewer’s time, effort, and constructive feedback. We hope that our responses have adequately addressed the reviewer’s concerns, and we would be happy to provide any further clarification if needed.

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

The TRACER proposed in this paper combines fine-grained evidence and logical evidence to address various image forgery detection tasks, including traditional manipulation, generative editing, and fully generated images. By constructing bidirectional sharing, interaction between specific levels, and progressive training, the model is able to output anomaly masks, authenticity assessments, and corresponding explanations. The core components of the model were evaluated through experiments and ablation analysis.

Critical strengths
  1. The evidence-centered formulation integrates forgery localization, authenticity judgment, and explanation into a unified framework, suitable for various forgery scenarios. This has made a meaningful contribution to the standardization of forgery analysis.

  2. This paper has a well-motivated design concept and empirical support. The bidirectional sharing mechanism associates spatial details with contextual reasoning. The ablation experiment confirmed its complementary contribution in localization and judgment.

Critical weaknesses
  1. The unseen-generator evaluation reports only Acc. Since localization is a central task, mIoU results would help assess whether the model also generalizes in locating anomalies. Therefore, I believe it would be best to use existing datasets to supplement the corresponding experiments.

  2. The method’s mIoU/F1 results are not consistently better than existing methods in this setting(such as Table 1). Please discuss these cases and clarify the benefits and limitations of the approach for fully generated images.

Questions for authors

For EQS, the appendix provides the criteria and prompts, but could the authors explain the rubric’s source and whether its reliability was checked through human ratings?

5 comments
AuthorsOpenReview ↗

Response to Reviewer jbQv (Part 1)

We sincerely thank you for recognizing our work and for your valuable suggestions. Below are our responses to your questions and concerns.

Response to W1:

Thank you for your valuable suggestion. Following your suggestion, we added forgery evidence localization experiments on four benchmarks with anomaly region annotations. SafeIMG [1] and ArtiBench [2] are AI-generated image benchmarks, while GeneVA [3] and ViF-CoT-4K [4] are AI-generated video benchmarks. These benchmarks include generators unseen during model training, including GPT Image 2, Qwen-Image, and Nano Banana. For a fair comparison, all baselines use the same training data as TRACER. The table below reports mIoU and F1 on each benchmark.

MethodSafeIMG [1]ArtiBench [2]GeneVA [3]ViF-CoT-4K [4]Average
mIoU / F1mIoU / F1mIoU / F1mIoU / F1mIoU / F1
DetectiveSAM.466 / .108.404 / .133.425 / .342.438‾\underline{.438} / .347.433 / .232
TruFor.468 / .124.460 / .193‾\underline{.193}.485 / .345.429 / .352.461 / .253
CAT-Net v2.470‾\underline{.470} / .111.486 / .156.474 / .371.422 / .372.463 / .252
LEGION(8B).468 / .115.494 / .168.494‾\underline{.494} / .376‾\underline{.376}.435 / .389‾\underline{.389}.473‾\underline{.473} / .262‾\underline{.262}
SIDA(13B).407 / .108.503 / .185.475 / .327.419 / .356.451 / .244
TRACER(4B).487 / .120‾\underline{.120}.499‾\underline{.499} / .215.503 / .394.447 / .396.484 / .281

TRACER achieves the highest average mIoU (0.484) and F1 (0.281) among all compared methods. These localization results provide further experimental support for TRACER's zero-shot generalization ability.

We have added Table 14 and the corresponding analysis under "Zero-Shot Generalization of Evidence Localization" in Appendix A.2 of the revised manuscript.

[1] Y.-Z. Wang, Y. Xiao, L. Yue, et al. AI-generated images challenge visual trust in high-risk scenarios. arXiv preprint arXiv:2607.22745, 2026.

[2] J. Park, M. Ahn, M. Kim, et al. See and fix the flaws: Enabling VLMs and diffusion models to comprehend visual artifacts via agentic data synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 35810–35820, 2026.

[3] J. Kang, M. B. Silva, P. Sangkloy, et al. GeneVA: A dataset of human annotations for generative text to video artifacts. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp. 6174–6183, 2026.

[4] Y. Li, W. Zheng, Y. Zhang, et al. Skyra: AI-generated video detection via grounded artifact reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 4482–4493, 2026.

AuthorsOpenReview ↗

Response to Reviewer jbQv (Part 2)

Response to W2:

Thank you for pointing out the differences in forgery evidence localization results in Table 1. We believe this is mainly related to how anomaly regions are represented in different benchmarks and how the model responds to the extent of these regions.

Anomalies in fully AI-generated images may appear as local structural or semantic inconsistencies, and their spatial boundaries are not always clear. Accordingly, the three benchmarks use different annotation forms: SynthScars uses polygons to outline specific anomalous parts, LOKI uses bounding boxes to mark anomaly regions, and RichHF expands human clicks on anomaly locations into circular region annotations. These representations describe anomaly regions at different levels of granularity, which may affect the overlap between predicted masks and annotated regions.

To show the differences between model predictions and annotated regions more clearly, we added Figure 10 to the appendix. In these representative examples, TRACER can identify key anomalies in hands, local animal structures, and text, but the predicted regions do not fully match the extent of the annotated regions. This may be related to the polygon-based localization supervision used to train TRACER, which may encourage it to outline specific local anomaly evidence.

We therefore aim to evaluate anomaly evidence localization more comprehensively in terms of response distributions, pixel ranking, and region coverage. Drawing on the use of NSS for artifact heatmap evaluation in ImageDoctor [5] and AP for forgery localization in Noiseprint [6], we added NSS and pixel AP, and also report P@1% and R@10%. Specifically,

  • NSS measures the strength of the model's responses within annotated anomaly regions relative to the response distribution over the entire image.
  • AP evaluates whether the model ranks anomalous pixels above background pixels.
  • P@1% is the proportion of the highest-scoring 1% of image pixels that fall within annotated anomaly regions.
  • R@10% is the proportion of all annotated anomaly pixels covered by the highest-scoring 10% of image pixels.
MethodLOKIRichHF
NSS / AP / P@1% / R@10%NSS / AP / P@1% / R@10%
MVSS-Net0.4983 / 31.07 / 40.71 / 24.121.0143 / 32.30 / 50.12 / 35.51
DetectiveSAM0.5046 / 31.25 / 41.75 / 23.520.9204 / 29.54 / 42.33 / 31.58
TruFor0.5245‾\underline{0.5245} / 32.90 / 39.33 / 23.030.9885 / 33.68‾\underline{33.68} / 51.88‾\underline{51.88} / 34.12
LEGION(8B)0.5082 / 33.04‾\underline{33.04} / 41.04 / 24.25‾\underline{24.25}0.8853 / 30.53 / 49.62 / 32.24
TRACER(4B)0.5373 / 33.48 / 41.67‾\underline{41.67} / 24.411.0124‾\underline{1.0124} / 33.87 / 53.08 / 35.05‾\underline{35.05}

TRACER performs well across these metrics. These results show that TRACER can capture key local anomaly evidence in fully AI-generated images. Together with the localization performance on benchmarks with unseen generators in the additional experiments for W1 above, these results further support TRACER's ability to generalize in capturing anomaly evidence in fully generated images. Appendix Figure 11 also illustrates how textual analysis complements localization when anomaly cues are subtle or lack clear spatial boundaries, providing additional context for authenticity assessment.

We have added Table 15, the corresponding analysis, and the calculation formulas under "Annotation Granularity and Localization Evaluation" in Appendix A.2 of the revised manuscript.

[5] Y. Guo, J. Liu, Z. Wang, et al. ImageDoctor: Diagnosing text-to-image generation via grounded image reasoning. In International Conference on Learning Representations, 2026.

[6] D. Cozzolino and L. Verdoliva. Noiseprint: A CNN-based camera model fingerprint. IEEE Transactions on Information Forensics and Security, vol. 15, pp. 144–159, 2020.

AuthorsOpenReview ↗

Response to Reviewer jbQv (Part 3)

Response to Q1:

Thank you for your question. Here, we explain the basis for the EQS design and provide additional evaluations by other multimodal large language model judges and human raters.

EQS evaluates the quality of anomaly analysis. Its design draws on related work that emphasizes explanation correctness, evidence relevance, and fine-grained analysis. AIGI-Holmes [7] uses an MLLM to score dimensions such as relevance, accuracy, and fine-grained detail, while FakeShield [8] emphasizes describing specific regions and explaining the basis for a judgment. Combining these ideas with the needs of our task, we define four EQS dimensions, with a total score of 0–10:

  • Visual evidence (0–4): Whether the described details and anomalies are specific, visible, and accurate.
  • Forensic reasoning (0–3): Whether the stated evidence supports the authenticity judgment.
  • Spatial grounding (0–2): For forged images, whether the explanation accurately describes where the relevant anomalies appear; for real images, whether the observations in the explanation match the actual image content.
  • Calibration (0–1): Whether the explanation avoids unsupported certainty and stays grounded in visible evidence.

Visual evidence receives the highest weight to prioritize accurate descriptions and evidence-supported authenticity judgments. For real images, the rubric encourages accurate observations and evidence-based analysis. It does not require the explanation to identify anomaly locations or invent forgery artifacts, nor does it treat visual inspection alone as proof of authenticity.

To reduce reliance on a single MLLM judge, we added DeepSeek V4.1 Flash and Kimi K3.1 Pro as MLLM judges. We also randomly sampled 50 images from each of four benchmarks covering different forgery scenarios and real images, and invited 5 volunteers to conduct independent blind evaluations using the same EQS rubric. The table below reports EQS scores (0–10).

MethodColumbiaOpenFor.SynthScarsSITD
DeepSeek / Kimi / HumanDeepSeek / Kimi / HumanDeepSeek / Kimi / HumanDeepSeek / Kimi / Human
GPT-5.56.92 / 7.23 / 8.045.08 / 6.04 / 4.246.77 / 7.29 / 7.478.16‾\underline{8.16} / 8.13 / 8.58
Gemini 3.1 Pro7.50 / 8.01‾\underline{8.01} / 8.585.82 / 5.77 / 4.767.05‾\underline{7.05} / 7.85 / 8.108.08 / 8.35‾\underline{8.35} / 8.69‾\underline{8.69}
Qwen3.7 Max7.92‾\underline{7.92} / 7.98 / 8.81‾\underline{8.81}6.23 / 6.15 / 5.406.00 / 7.38 / 7.137.84 / 7.08 / 8.13
FakeShield(22B)5.08 / 5.69 / 4.846.77 / 7.46‾\underline{7.46} / 6.05— / — / —6.18 / 6.42 / 7.41
FakeVLM(7B)— / — / —7.54‾\underline{7.54} / 7.31 / 8.44‾\underline{8.44}6.77 / 6.08 / 8.476.78 / 6.72 / 7.88
SIDA(13B)4.09 / 5.92 / 4.515.54 / 5.77 / 3.924.85 / 4.13 / 3.415.28 / 5.71 / 6.70
TRACER(4B)8.42 / 8.16 / 9.128.13 / 7.95 / 8.977.85 / 7.74‾\underline{7.74} / 8.32‾\underline{8.32}8.65 / 8.47 / 8.86

As shown in the table, the MLLM judges and human raters show broadly consistent overall ranking trends across methods, lending support to the reliability of the EQS rubric across evaluators. Under this shared rubric, TRACER consistently ranks among the top methods in both MLLM-based and human evaluations, further supporting the quality of its anomaly analysis.

We have added Table 13 and the corresponding analysis under "Explanation Quality Across MLLM and Human Judges" in Appendix A.2 of the revised manuscript.

[7] Z. Zhou, Y. Luo, Y. Wu, et al. AIGI-Holmes: Towards explainable and generalizable AI-generated image detection via multimodal large language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 18746–18758, 2025.

[8] Z. Xu, X. Zhang, R. Li, et al. FakeShield: Explainable image forgery detection and localization via multi-modal large language models. In International Conference on Learning Representations, 2025.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank all reviewers and the Area Chair for their time and constructive feedback on TRACER. We appreciate the reviewers' recognition of the unified formulation, complementary evidence modeling, careful training design, and empirical results. For example:

  • The evidence-centered formulation integrates forgery localization, authenticity judgment, and explanation into a unified framework, suitable for various forgery scenarios. (Reviewer jbQv)
  • The dual-evidence modeling paradigm is intuitive and well-motivated. It complements low-level forensic traces and high-level semantic logical constraints, with bidirectional cross-branch information exchange to boost localization and classification. (Reviewer cxoP; also recognized by Reviewer jbQv)
  • The authors carefully design interaction strategies and progressive training scheduling to mitigate distribution mismatch between two pre-trained backbones, not only proposing an architecture but also investigating training stability. (Reviewer cxoP)
  • The generalizability experiments demonstrate promising performance on unseen datasets and manipulation types. Compared with previous LLM-based image forgery localization methods, the proposed method uses only 4B parameters, demonstrating its parameter efficiency. (Reviewer UBQX)

We also thank the reviewers for their constructive suggestions for improving this work. We have provided point-by-point responses and revised the manuscript. The main revisions are highlighted in RED and summarized below:

  • BES and Fusion Baselines. We have clarified the design of BES and progressive training, and compared BES with feature concatenation and standard cross-attention using the same backbones and training procedure (Sec. 4.2, Table 5). (reviewers cxoP, UBQX)
  • Generalization and Robustness. We added localization experiments on image and video benchmarks with unseen generators, extended robustness testing to ArtiBench, and compared robustness across methods (Appendix A.2, Tables 10 and 14, Fig. 8). (reviewers cxoP, jbQv)
  • Explanation Quality. We have clarified the EQS scoring criteria and added evaluations by two additional MLLM judges and five independent human raters using the same rubric (Appendix A.2, Table 13). (reviewers cxoP, jbQv)
  • Localization and Annotation Granularity. We have analyzed how annotation differences affect localization results and added representative visualizations and supplementary metrics, including NSS, pixel AP, P@1%, and R@10% (Appendix A.2, Table 15, Fig. 10). (reviewers cxoP, jbQv)
  • Motivation and Dual-Evidence Modeling. We have clarified that fine-grained and logical evidence represent two key perspectives for identifying forgery cues, rather than an exhaustive classification of anomalies. We added statistical analysis and visualizations of the two streams to support this motivation (Sec. 4.3, Fig. 5; Appendix A.2, Fig. 9). (reviewer UBQX)

We hope these clarifications and additional results address the reviewers' concerns. We welcome further comments and questions during the discussion. Thank you again for your valuable feedback.

AuthorsOpenReview ↗

Final General Response

We sincerely thank all three reviewers and the Area Chair for their time and constructive feedback. We appreciate the reviewers’ recognition of our unified formulation, well-motivated dual-evidence modeling, careful training design, and promising generalization performance. Our main revisions and additional results are summarized below:

  • Motivation and Unified Design. We clarified how joint modeling of fine-grained and logical evidence supports detection across diverse forgery scenarios. Added statistical analyses and stream visualizations support this motivation (Sec. 4.3, Fig. 5; Appendix A.2, Fig. 9).
  • Controlled Comparisons. We added fusion baselines and a parameter-scale ablation. BES achieves the best overall performance, and the 2B variant outperforms the 4B fusion baselines overall, supporting the effectiveness of our interaction design (Sec. 4.2, Table 5).
  • Generalization and Robustness. We added unseen-generator localization experiments, extended robustness evaluations, and compared methods under common perturbations. TRACER achieves the highest average localization performance on the four added benchmarks (Appendix A.2, Tables 10 and 14, Fig. 8).
  • Explanation Quality. We clarified the EQS rubric and added evaluations by two additional MLLM judges and five independent human raters. Their broadly consistent results further support TRACER’s explanation quality (Appendix A.2, Table 13).
  • Localization and Annotation Granularity. We added supplementary localization metrics and visualizations to analyze annotation differences and assess the model’s ability to identify key anomaly evidence (Appendix A.2, Table 15, Fig. 10).

We are encouraged that all three reviewers now give positive ratings (3: Weak acceptance), with UBQX raising their rating after reviewing our responses and additional experiments. We appreciate the constructive discussion and will continue to improve our work.

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

Related papers

Discussion (0)

Sign in to comment.