- 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
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:
- Final-layer Feature Fusion concatenates each stream's final-layer features with MLP-projected features from the other stream.
- Layer-wise Feature Fusion uses the same interaction as the first baseline, but applies Hybrid Layer Interaction to exchange information at multiple depths.
- Layer-wise Cross-Attn uses each stream's own
Qto attend only to the other stream'sK/Vat 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:
| Setting | CASIAv1+ | CocoGlide | SynthScars | SITD |
|---|---|---|---|---|
| mIoU / Acc | mIoU / Acc | mIoU / Acc | Acc | |
| Final-layer Feature Fusion (4B) | .848 / .866 | .605 / .691 | .476 / .898 | .948 |
| Layer-wise Feature Fusion (4B) | .880 / .891 | .605 / .727 | .550 / | .950 |
| Layer-wise Cross-Attn (4B) | .590 / .460 | .485 / .428 | .471 / .260 | .922 |
| Layer-wise BES (2B) | / | / | / .999 | |
| 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.
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:
| Benchmark | Metric | No Perturbation | JPEG Compression QF=90 | JPEG Compression QF=80 | JPEG Compression QF=70 | Gaussian Noise σ=0.02 | Gaussian Noise σ=0.06 | Gaussian Noise σ=0.10 | Gaussian Blur Ksize=3 | Gaussian Blur Ksize=5 | Gaussian Blur Ksize=7 | Gamma Correction γ=0.7 | Gamma 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 |
| SynthScars | mIoU | .560 | .543 | .545 | .544 | .547 | .534 | .517 | .557 | .560 | .559 | .551 | .550 |
| SynthScars | Acc | .999 | 1.000 | .999 | .996 | .962 | .800 | .722 | 1.000 | 1.000 | 1.000 | .988 | .999 |
| ArtiBench | mIoU | .499 | .486 | .485 | .485 | .478 | .473 | .471 | .493 | .498 | .501 | .480 | .483 |
| ArtiBench | Acc | .948 | .960 | .972 | .962 | .808 | .526 | .408 | .990 | .994 | .996 | .922 | .960 |
| COCO | Acc | .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.
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
Qto query another’sK/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'sK/Vinto 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 ownK/Vand the projected cross-streamK/V, allocating attention between them based on its currentQ. 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:
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: