acceptodds
Under review as a conference paper at ICLR 2027

DeCue: Learning Generator-Invariant Forensic Representations via Feature Decoupling for AI-Generated Image Detection

Abstract

The rapid growth of image generative models, from GANs to diffusion models, has enabled the widespread creation of photorealistic synthetic images, posing serious threats to digital media authenticity and information integrity. A fundamental bottleneck of AI-generated image detection is the poor cross-generator generalization of existing detectors. We attribute this generalization failure to the inherent entanglement of deep representations. Detectors tend to overfit to superficial semantic content features rather than learning intrinsic generation-invariant forensic traces that transfer across architectures. To address this issue, we present DeCue, a novel feature decoupling framework that explicitly disentangles entangled representations into forensic cue features encoding generation-specific discriminative artifacts, and content features that capture image semantic information. The proposed DeCue achieves robust feature decoupling through hard fake sample construction and cross-reconstruction constraint. The hard fake samples are strategically synthesized to eliminate trivial content-level disparities between real and fake images, compelling the model to learn generalizable forensic cues instead of relying on content shortcuts. The cross-reconstruction constraint further provides explicit supervisory signals to guaranty effective feature decoupling. Extensive evaluations across 8 benchmarks, including 5 standard academic benchmarks and 3 challenging in-the-wild benchmarks, show that our DeCue achieves a mean accuracy of \textbf{99.23\\%}, substantially outperforming the state-of-the-art detector by \textbf{11.29\\%}, demonstrating its strong robustness and superior cross-generator generalization capability in practical forensic scenarios.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
AI-generated image detection, feature decoupling, single-class classification

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_iqEKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper attributes the poor cross-generator generalization of existing AIGI detectors to the entanglement of semantic content and forensic cues in deep representations, which causes detectors to overfit to semantic shortcuts. To address this issue, it proposes DeCue, which consists of two components: hard fake sample construction and feature decoupling pipeline. DeCue reports an average accuracy of 99.23% across eight benchmarks, including five academic benchmarks and three in-the-wild benchmarks, outperforming the SOTA baseline by 11.29 percentage points.

Critical strengths

1.The experiments are comprehensive. The paper evaluates the proposed method on eight benchmarks, compares it with 16 competing methods, and provides visualizations, ablation studies, and robustness evaluations.

2.The paper generates semantically aligned fake samples from real images, designs a decoupling module to separate forensic cues from semantic features, and trains the model using the fused representation. This is a meaningful direction with clear motivation.

3.The decoupling losses are reasonably designed, and the visualization results provide evidence of their effectiveness.

Critical weaknesses

1.The values in the “Mean” column of Table 1 do not appear to be arithmetic averages. The authors should verify whether these values were calculated correctly.

2.The full model achieves 98.20% on AIGI-Now in Table 2, whereas Table 1 reports 98.90%. Similarly, the full model achieves 94.10% on AIGIBench in Table 3, whereas Table 1 reports 94.80%.

3.Around line 222, “x = Exr” should be “z = Exr.” In addition, panels (d) and (e) are mislabeled in the caption of Figure 4.

4.The hard fake sample generation procedure largely follows DDA, while feature decoupling has already been extensively explored in both deepfake detection and AIGI detection. Therefore, the methodological novelty is limited.

5.The t-SNE and Grad-CAM visualizations provide only limited evidence for successful feature decoupling. The paper lacks quantitative evaluations or analytical tests to substantiate this claim.

Questions for authors

See my comments above.

Reviewer_thDLOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DeCue, an AI-generated image detection method for cross-generator generalization, aiming to address the problem that existing detectors over-rely on image semantic content and struggle to learn general forensic cues. The method first constructs semantically consistent, difficult-to-forge samples through latent space perturbation reconstruction and data mixing to reduce the impact of content shortcuts on the detector. Then, through cross-reconstruction constraints and content alignment constraints, deep features are explicitly decoupled into forensic features and content features. Finally, forensic features and feature reconstruction residuals are fused to achieve image detection with strong cross-generator generalization capabilities.

Critical strengths

None

Critical weaknesses
  1. The construction process of hard fake-samples largely borrows from the DDA approach, while feature decoupling has been extensively studied, with many methods already applied to AI-generated image detection. Therefore, the overall innovation of this paper is relatively limited.
  2. This paper exhibits a strong engineering style, lacking a clear core guiding principle. Each module seems to be designed specifically for a particular problem, making the entire framework more like a simple stacking of existing modules, lacking theoretical motivation and design logic.
  3. Why is it still necessary to fuse feature-level reconstruction residuals for final discrimination after obtaining the evidence-gathering features through decoupling? The paper explains this by stating that "relying solely on evidence-gathering features may lack robustness in complex post-processing scenarios," but this reason is unconvincing. The purpose of feature decoupling is precisely to reduce biases caused by non-evidence-gathering factors such as content. If the alignment of hrh_r and hah_a primarily focuses on semantic consistency, are post-processing-related pseudo-correlation features still likely to be retained in the evidence-gathering features? Furthermore, will introducing additional reconstruction residuals reintroduce these interfering factors, thereby weakening the model's generalization ability?
  4. The proposed method in Table 2 achieves an mAcc of 98.20% on AIGI-Now, while the result reported in Appendix Table 14 is 98.90%. Please explain why the same method yields different experimental results on the same dataset.
  5. There is an inconsistency between the titles and captions of subfigures (d) and (e) in Figure 4. Assuming (d) shows the decoupled content features, its visualization still demonstrates clear true/false separability. This implies that content features may still retain strong true/false discrimination information. It is recommended that the authors supplement experiments using only content features for classification to verify whether the proposed method truly achieves effective feature decoupling.
  6. There is a lack of comparative experiments with different backbones. Furthermore, Table 2 shows that using only Vanilla DINOv3 achieves an mAcc of 95.86%, even surpassing the performance of all the comparative methods in Table 14. This phenomenon warrants further explanation.
  7. The proposed method achieves near 100% performance on almost all datasets, a result so close to perfection that it raises questions about the reliability of the experiments. Although the paper provides an anonymous code repository, it is hoped that the authors will clarify whether the complete training code and data processing flow have been made public to support the independent reproduction of the experimental results.
Questions for authors
  1. Besides the construction of hard forgery samples based on DDA and existing feature decoupling methods, what are the key methodological contributions of this paper?
  2. If the proposed framework can successfully separate forensic clues from content features, why is it necessary to introduce reconstruction residuals for classification? Will these residuals reintroduce false relevance and weaken generalization ability?
  3. Figure 4 shows that the decoupled content features still contain a significant amount of information distinguishing between true and false data. Can the authors provide content-based classification results to demonstrate the effectiveness of the claimed decoupling method?
  4. Why is the mean accuracy (mAcc) of Table 2 98.20% on the AIGI-Now dataset, while Table 14 is 98.90%?
  5. The original DINOv3 baseline model has achieved an mAcc of 95.86%, outperforming all the comparison methods in Table 14. How much of this performance improvement can be attributed to the proposed method, rather than a more powerful backbone network or training scheme?
  6. Given that the method demonstrates near-perfect performance on several challenging benchmarks, could the authors explain the measures taken to ensure fair comparisons, prevent data leakage, and verify the reproducibility of the results?
Reviewer_mHvfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DeCue studies cross-generator AI-generated image detection by separating visual representations into forensic cue and semantic content branches. It constructs groups of content-aligned hard negatives from MSCOCO real images through Stable Diffusion v2.1 VAE reconstruction, latent perturbations, and pixel/frequency alignment. A LoRA-adapted DINOv3 backbone is regularized against a frozen anchor. Cross-reconstruction and content-alignment objectives supervise the branches; cue features and feature reconstruction residuals are fused through cross-attention and classified using a KAN. Evaluation spans eight benchmarks covering different generator families and in-the-wild images, with ablations and robustness tests. The reported results are promising, but the methodological novelty, the connection between the motivation and the chosen backbone, and the consistency and attribution of experimental gains require clarification.

Critical strengths
  1. The paper addresses an important practical problem and evaluates beyond a single generator family, including commercial systems and in-the-wild images. Training uses MSCOCO real images and reconstructed negatives rather than the benchmark training splits. If confirmed under a clearly specified protocol, the reported Chameleon accuracy of 99.94% and AIGIBench accuracy of 94.80% would be consequential. The appendix also reports strong AP, so the reported advantage is not solely a fixed-threshold accuracy effect.
  2. The framework combines a reasonable strategy for reducing content shortcuts with feature-level supervision and residual fusion. The ablations provide preliminary evidence of component usefulness: removing cross-reconstruction changes AIGI-Now mAcc from 98.20% to 97.63%, and removing residual features changes AIGIBench mAcc from 94.10% to 87.85%. These results motivate further investigation, although they do not by themselves establish novelty or isolate the mechanism.
Critical weaknesses
  1. The methodological novelty is not sufficiently distinguished from prior work. Content-aligned reconstructed negatives have clear precedents in AlignedForensics (Rajan et al., ICLR 2025) and DDA (Chen et al., NeurIPS 2025); DeCue explicitly builds on DDA. Separating content from forgery information and using cross-reconstruction also have precedents in Exploring Disentangled Content Information for Face Forgery Detection (Liang et al., ECCV 2022) and UCF (Yan et al., ICCV 2023). These face-forgery methods differ in task and reconstruction target, so this does not establish equivalence of their objectives to DeCue's feature-space loss. Nevertheless, the distinction needs explanation. Moreover, OmniAID (Guo et al., 2025) already combines a decoupled semantic/artifact expert architecture with semantically aligned VAE-reconstructed training pairs, making Section 2.2's broad claim about prior methods not combining sample-level alignment and feature-level decoupling insufficiently qualified. The contribution should be located more precisely in the perturbed sample groups, the groupwise feature-reconstruction objective, or their interaction with residual fusion, with evidence that the claimed increment matters.
  2. Figure 2 uses CLIP to motivate the proposed feature-decoupling framework, whereas the method adopts DINOv3 as its backbone. While the paper provides a general rationale for choosing DINOv3, it remains unclear why DINOv3 is preferred over CLIP for this framework and whether applying the same design to CLIP would also be effective.
  3. I am concerned about the consistency of the experimental results and the attribution of gains to decoupling. The eight displayed accuracies in Table 1 yield unweighted means of 98.89875% for DeCue and 84.9225% for GAPL, rather than the reported 99.23% and 87.94%; no alternative weighting is specified. Full-model AIGI-Now is 98.20% in Table 2 versus 98.90% in Table 1, and AIGIBench is 94.10% in Table 3 versus 94.80% in Table 1. I would like to understand what causes these differences in the reported results.
Questions for authors

See Critical Weaknesses.

Comments on the paper

1 comment
Area_Chair_NUV3OpenReview ↗

Reviewer's responses to authors

Dear reviewers,

If not done already, could you please take a look and reply briefly to the "Official Comment(s) by Authors" / Rebuttals, possibly by an acknowledgement that you have seen it.

Feel free to respond in more detail, in particular, if you have further questions or opinions to the replies from the authors.

With best regards, an AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.