acceptodds
Under review as a conference paper at ICLR 2027

Learning Continuous Source Responses For Generalizable AI-Generated Image Detection

Abstract

Advances in image generation have made synthetic images increasingly difficult to distinguish from real photographs, raising concerns about the trustworthiness of visual media. Existing AI-generated image detectors often perform well on in-domain data, but their robustness and cross-generator generalization remain limited. These limitations are commonly attributed to overfitting to shortcut cues. Although many methods seek to suppress shortcut learning, most retain binary classification as the training task without reconsidering how the task itself shapes the learned representations. We introduce CuRe, a framework for learning Continuous Source Responses that revisits authenticity detection from the perspective of the training task. CuRe reformulates backbone adaptation as regression of real-generated mixing ratios, providing finer supervision that encourages the model to capture authenticity-related variation beyond binary endpoint separation. We further select a compact source-response subspace to suppress nuisance variation and limit the final classifier's access to potential shortcut cues. Across ten public benchmarks, CuRe achieves an average balanced accuracy of 89.7%, exceeding the second-best method by 5.2%. Further experiments demonstrate consistent generalization gains across visual backbones and strong robustness to common image degradations. Code is available at https://anonymous.4open.science/r/CuRe-0CF7/

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI-generated image detection, Task Reformulation

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

This paper proposes the CuRe framework, which reformulates binary classification into regression for real-generated mixing ratio estimation. It further selects a source-response subspace to suppress nuisance variation and restrict the final classifier's access to potential shortcut cues. The generalization and robustness are validated across 10 benchmarks.

Critical strengths

1、The proposed method achieves consistent and improved forgery detection performance across 10 diverse benchmarks, showing strong generalization against unseen generative models. 2、The proposed paradigm converts the conventional discrete AIG detection task into a continuous learning formulation for mixing ratio estimation. The method design aligns well with the core research motivation.

Critical weaknesses

Main Weakness:

  1. Insufficient justification for the proposed method over simpler shortcut mitigation strategies.

Figure 2 highlights the separation of features associated with different JPEG quality levels (q95 vs. q50) as evidence of format-related shortcut learning. However, such shortcuts can be mitigated through straightforward data augmentation or bias-controlled training, such as randomizing JPEG quality levels independently of real/fake labels or balancing compression distributions across the two classes. Moreover, the authors already apply JPEG format alignment during training. It remains unclear why these simpler strategies are insufficient and what additional benefits CuRe provides beyond existing shortcut mitigation techniques. A controlled comparison with these baselines is necessary to better justify the motivation and necessity of the proposed framework.

  1. Unclear relationship between semantic representations and forgery detection performance.

The paper argues that excessive reliance on semantic information contributes to shortcut learning, while Figure 2 suggests that suppressing semantic grouping leads to better separation between real and generated images. However, the relationship between semantic discrimination and forgery detection performance is not clearly established. Does stronger semantic discrimination necessarily impair authenticity detection, and is there quantitative evidence supporting this hypothesis? Furthermore, if reducing semantic dependence is central to the proposed approach, it is unclear why the framework relies on semantically powerful pretrained vision encoders with limited LoRA adaptation. Why not train a task-specific detector from scratch or adopt a backbone with weaker semantic priors? These choices require stronger motivation and controlled experimental validation.

Minor Weakness:

1、Experimental results show more significant improvements on in-the-wild benchmarks, yet the underlying reasons are not analyzed. Moreover, the definition of "in-the-wild" is not explicitly provided, and details such as dataset split protocols are missing.

Questions for authors

Why is the subspace dimension set to 128? There is no experimental evidence proving this dimension is optimal.

Actually, I am leaning toward a borderline rating for this paper. However, I would be willing to raise my score if the authors can adequately address my first two major weaknesses.

3 comments
AuthorsOpenReview ↗

Augmentation baselines, semantic representations, in-the-wild definition, and sensitivity to k

Thank you for the thoughtful questions, which have helped us improve the paper.

W1 (Major): Why are simpler shortcut-mitigation strategies insufficient?

Response: We agree that data augmentation and bias-controlled training can mitigate specific shortcuts. However, they require knowing in advance which bias to remove, and a dedicated augmentation has to be designed for each one. For a controlled comparison, we added JPEG, blur, and resizing augmentations to the binary baseline, keeping the backbone, training data, and base training setup the same, and compared the three models under different perturbations:

Balanced accuracy (%) on AIGIBench.

MethodCleanJPEG q=30Resize s=0.5Resize s=1.5Blur σ=2.0Worst caseAverage
Binary + JPEG80.279.475.274.168.668.675.5
Binary + JPEG, blur, and resize79.575.378.178.670.370.376.4
CuRe92.890.492.292.990.490.491.7

Overall, targeted augmentation brings only limited improvement: with augmentations added, the binary baseline is still more than 9 points below its clean accuracy in the worst case, and its clean accuracy does not improve either. In contrast, CuRe drops by at most 2.4 points under these perturbations.

In addition, generator-specific shortcuts cannot necessarily be resolved by adding more generators to the training set, since a new generator may differ in distribution from those already seen. Compared with approaches that rely on expanding the training data, actively suppressing shortcuts is therefore more effective.

W2 (Major): Semantic representations and the choice of pretrained backbone

Response: Thank you for pointing this out. The wording in the paper was indeed imprecise. We have corrected the relevant statements in the introduction and the caption of Figure 2 in the revised PDF (marked in blue).

We do not believe that semantic discrimination ability impairs detection. The finer visual details captured by pretrained representations help distinguish real from generated images. What we want to avoid is decisions that rely on coarse differences in category or content, which are unrelated to authenticity and can easily become shortcuts. This is also what Figure 2 is meant to show. Such shortcuts form easily in practice, since real and generated images in common training sets often differ in content.

We further ran an experiment: removing the suppression of content directions in Stage 2 (λ_s = 0) reduces the Overall balanced accuracy from 89.7 to 80.3. This indicates that keeping the decision away from directions dominated by content variation is important for generalization.

We still use a semantically strong pretrained backbone because we want to preserve these visual details, which is why we adapt it only with LoRA. Pretrained representations are useful in themselves: in Table 8, stronger pretrained backbones give better results for both binary training and CuRe, whereas detectors trained on the same data without such pretrained representations (NPR and SAFE) reach only 56.8 and 62.4 Overall balanced accuracy. The key is therefore not whether to use a semantically strong representation, but how to adapt it.

Minor: Definition of in-the-wild, the reason for larger gains, and the data split

Response: In-the-wild benchmarks contain images collected from real online environments such as social media, or images that simulate the degradations of online sharing. Their generators are more diverse and often unknown, and the images are frequently compressed or resized. CuRe generalizes well to unseen generators and is robust to postprocessing, which is why its gains are larger there.

Regarding the data split, CuRe is trained only on the training set of GenImage SD v1.4, epochs are selected on its official validation set, and all 10 benchmarks are used only for testing. We have added these details to the revised PDF.

Q: Why is the subspace dimension set to 128?

Response: We added a sensitivity analysis on k, using the same Stage 1 checkpoint and only recomputing the subspace and retraining the linear classifier.

Balanced accuracy and AUC (%) on AIGIBench.

k326496128 (default)160192224
B.Acc91.791.893.092.892.191.091.2
AUC96.095.297.597.596.895.795.3

The results are not very sensitive to k: for k between 32 and 224, the balanced accuracy stays between 91.0 and 93.0, which suggests that most of the signal is concentrated in a small number of dimensions. When k becomes too large (192, 224), performance drops and gradually approaches that of using the full features.

We should note that we had not previously run a broad search over k. We had only tried a few values. We have added this sensitivity analysis to the revised PDF (Appendix B.5).

We hope this addresses your concerns, and we are happy to discuss further.

Reviewer_MSWuOpenReview ↗

Thanks for the response

Thank you for the response, which has addressed my concerns. I am willing to raise my score.

Dear Reviewer MSWu,

We are grateful for your quick follow-up and for your willingness to re-evaluate our submission in light of our response. We truly appreciate your positive assessment of our work.

If anything remains unclear or you have further comments, we would be happy to address them, and we will keep following this thread through the end of the author-response period.

Thank you once more for your time and constructive feedback.

Best regards, The Authors

Reviewer_zgBvOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes CuRe, a two-stage framework for generalizable AI-generated image detection. Instead of conventional binary supervision, the method trains a visual backbone to predict continuous mixing ratios between real and generated images. It then constructs a compact source-response subspace to suppress nuisance variations and improve generalization. Experiments across ten benchmarks demonstrate strong performance.

Critical strengths
  1. Controlled evidence that the training objective matters. Under a matched LoRA protocol with repeated seeds, replacing binary supervision with ratio regression gives a clear gain. The gain holds consistently across four different backbones and several LoRA ranks. This matters to my assessment because it is the main support for the paper’s central claim. It shows that the improvement is not tied to one backbone or configuration, which makes the idea likely to transfer to other detectors.

  2. Strong robustness to common post-processing. CuRe stays nearly flat under JPEG compression, resizing, and blur, while competing methods degrade substantially. This matters because real-world detection almost always involves degraded images. Robustness of this kind is practically more valuable than small gains on clean benchmarks, and it is the strongest reason the method could be useful even if its novelty relative to Mixup turns out to be limited.

Critical weaknesses
  1. Unclear mechanism of continuous source supervision. Intermediate images are formed by pixel-space interpolation between randomly paired real and generated images. Such mixtures do not necessarily correspond to meaningful intermediate levels of authenticity, and the model may partly learn to detect blending itself rather than transferable forensic cues. Same-source mixtures appear only in Stage 2, not in Stage-1 training. To support the proposed explanation, the authors should report how the Stage-1 model scores real-real and generated-generated mixtures, and whether the gain persists when such mixtures are added to training with constant targets.

  2. Limited novelty and unclear role of the subspace. Regressing the mixing ratio between two classes is essentially inter-class Mixup, which is neither cited nor compared against. For Stage 2, the improvement over the direct Stage-1 score appears mainly in balanced accuracy, with little change in AUC. This suggests that the subspace largely corrects the decision threshold rather than isolating authenticity-related information. Comparisons with standard Mixup, simple score calibration, and LDA are needed to establish the contribution of each component.

Questions for authors
  1. How does CuRe compare with standard Mixup, inter-class-only Mixup, and label smoothing under the same backbone and training protocol?
  2. What does the Stage-1 model predict on real-real and generated-generated mixtures? Does the gain persist when such same-source mixtures are added to Stage-1 training with constant targets?
  3. Can a calibrated threshold or a one-dimensional classifier on the Stage-1 score match the full method? How does the proposed subspace compare with LDA?
  4. How large is the advantage over strong baselines such as DDA or B-Free when they are reproduced on the same PE Core backbone?
5 comments
AuthorsOpenReview ↗

Comparison with Mixup and label smoothing; is Stage 1 detecting mixing? (1/2)

Thank you for the careful and constructive comments. We address each point below.

Q1: Comparison with standard Mixup, inter-class Mixup, and label smoothing

Response: We agree that Stage 1 is, in form, inter-class Mixup with λ ~ U(0, 1). We have cited Mixup and discussed this relation in the revised PDF (Section 3.1, marked in blue). The table below compares the training objectives under the same setting, reporting balanced accuracy (%) on AIGIBench and on all 10 benchmarks.

Training objectiveInputLabelAIGIBench B.AccOverall B.Acc (10 benchmarks)
BinaryUnmixedHard80.274.2
Label smoothing (ε=0.1)UnmixedSmoothed80.776.0
Standard Mixup (random pairing of two images)MixedMixing ratio81.575.7
Inter-class Mixup only (CuRe Stage 1)Mixed (real and generated only)Mixing ratio89.881.1
CuRe (Stage 1 + Stage 2)Mixed (real and generated only)Mixing ratio92.689.7

Overall, label smoothing and standard Mixup are both close to binary training, and only Stage 1, which supervises the mixing ratio between real and generated images, is clearly better (89.8 vs. 81.5 on AIGIBench). This shows that the gain comes neither from soft labels nor from the mixing operation itself, but from ratio supervision over the continuous transition between real and generated images, which makes the model more sensitive to source changes. This is consistent with the central claim of the paper.

Q2 / W1: Is Stage 1 detecting mixing?

Response: The table below reports the average prediction (the predicted proportion of the generated image) of two models on three types of mixing trajectories on AIGIBench: real to generated (RF), real to real (RR), and generated to generated (FF).

TrajectoryModelα=00.20.40.60.81.0
RFStage 10.1030.2870.4400.5260.6620.822
RFStandard Mixup0.0750.2340.3480.4240.5070.674
RRStage 10.1030.2740.4200.4190.2740.103
RRStandard Mixup0.0750.1900.2460.2460.1920.075
FFStage 10.8220.6770.5610.5610.6770.822
FFStandard Mixup0.6740.5630.5340.5340.5640.674

Stage 1 does respond to mixing within the same source: on the RR and FF trajectories, its predictions move toward the middle. However, it is not merely detecting mixing. On the RF trajectory, α = 0.2 and α = 0.8 have the same degree of mixing, yet the predictions are clearly different (0.29 vs. 0.66). A model that only detects mixing could not tell these two cases apart. Moreover, Stage 1 outperforms binary training by nearly 7 points on the 10 benchmarks, whose test images are never mixed.

Adding mixtures from the same source to training (Standard Mixup) reduces the response to mixing to some extent, but its detection performance is not much better than binary training. This may be because about half of the training samples then reduce to hard labels, leaving less supervision on intermediate ratios, so the model relies on shortcuts more easily (its predictions on generated images also shift toward real, e.g., 0.67 vs. 0.82 on unmixed generated images).

We believe the effect of mixing arises mainly at the image level: a mixed image deviates from natural images, which adds a response along extra feature directions. This response appears on both the RR and FF trajectories, so it reflects the mixing itself rather than the source. We therefore keep the full ratio supervision in Stage 1, and use Stage 2 to suppress the directions related to mixing.

AuthorsOpenReview ↗

Simpler readouts vs. Stage 2; comparison with DDA on PE Core (2/2)

We continue our response below with the remaining two questions.

Q3 / W2: Comparison with a calibrated threshold, a 1D classifier, and LDA

Response: The differences on AIGIBench are small, and results on a single benchmark may be affected by the characteristics of that dataset. We therefore also report results on all 10 benchmarks. The calibrated threshold and the 1D logistic regression are fitted only on the validation set of the training domain.

MethodAIGIBench B.AccOverall B.Acc (10 benchmarks)
Stage 1 direct prediction (threshold 0.5)89.881.1
Stage 1 + calibrated threshold90.782.4
Stage 1 + 1D logistic regression90.081.0
LDA88.379.1
Full feature + linear classifier89.779.4
CuRe (subspace)92.689.7

None of these simpler readouts matches the full method: the best of them reaches only 82.4 on the 10 benchmarks, compared with 89.7 for CuRe. A calibrated threshold or a 1D classifier only rescales the Stage 1 score. It cannot remove the mixing and content factors already contained in that score (see Q2). In addition, the validation and test distributions are not identical, so a threshold calibrated on the validation set transfers poorly.

LDA and the full feature linear classifier determine their directions directly on all features using real and generated labels, and likewise do not suppress these factors. The role of Stage 2 is therefore not calibration. It selects, in feature space, the directions that respond strongly to changes between real and generated images, and suppresses the directions related to mixing and content.

Q4: Advantage over DDA and B-Free on the same PE Core backbone

Response: Please see our response to Reviewer 1XZ2 (Q1), where we port strong prior methods to PE Core L/14. The backbone does matter: some methods improve clearly with PE Core (for example, UnivFD rises from 66.1 to 74.9 Overall). Among the ported methods, DDA is the best with 81.4 Overall, while CuRe reaches 89.7, so the advantage of CuRe remains.

For B-Free, the official release does not provide the training code. We were therefore unable to reproduce it faithfully on PE Core.

We hope the responses above address your concerns, and we are happy to discuss further.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer zgBv,

We sincerely appreciate your careful review and your recognition of our work. We would like to briefly follow up on our response. We added comparisons with Mixup and label smoothing, an analysis of Stage 1 on same-source mixtures, comparisons between Stage 2 and simpler readouts, and a comparison with DDA on the same PE Core backbone. We would be happy to discuss further if anything remains unclear.

Best regards, The Authors

Reviewer_1XZ2OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CuRe addresses generalizable AI-generated image detection by learning continuous source responses rather than relying solely on binary real-versus-generated supervision. It constructs pixel-space interpolations between real and generated images and adapts a pretrained visual encoder using LoRA to predict their mixing ratios. The adapted encoder is then frozen, and a compact source-response subspace is selected using real–generated, real–real, and generated–generated interpolation trajectories to retain source-sensitive directions while suppressing content and mixing-related variation.

Critical strengths
  1. An interesting reformulation of backbone adaptation. CuRe replaces binary endpoint supervision during backbone adaptation with continuous mixing-ratio prediction between real and generated images. This offers an alternative way to learn source-sensitive representations, complemented by a subspace-selection stage designed to suppress content and mixing-related variation.

  2. Broad evaluation covers generalization and robustness. The paper evaluates CuRe on ten public benchmarks spanning standard and in-the-wild settings. Additional experiments examine robustness to JPEG compression, resizing, and Gaussian blur, providing a broader assessment than clean-image detection alone.

Critical weaknesses

The results for CuRe and previous methods in Tables 1 and 2 are not directly comparable. The most significant difference is the use of different backbones.

The authors retrained NPR, UnivFD, FatFormer, and SAFE using the same paradigm as CuRe while retaining their original architectures, some based on ResNet and others on CLIP ViT-L/14. Among the methods that were not retrained, DDA uses DINOv2, while GAPL and DGS-Net use CLIP ViT-L/14. None of these methods uses the same PE Core L/14 backbone as CuRe. Figure 4 shows that PE Core L/14 performs significantly better than the other backbones. Therefore, the contribution of the CuRe framework itself, beyond the choice of backbone, remains unclear. The authors should include baseline methods that also use PE Core L/14 or retrain selected baselines with this backbone to enable a more controlled comparison.

In Appendix Table 8, the AUC of CuRe drops from 97.53 with PE Core L/14 to 80.13 with CLIP ViT-L/14. By comparison, GAPL, the best-performing previous method in Table 2, achieves an AUC of 96.5 using CLIP ViT-L/14. This observation further raises the question of whether the stronger results of CuRe in Tables 1 and 2 are primarily attributable to its backbone.

Questions for authors
  1. Backbone-controlled comparisons. Could you evaluate selected strong prior methods using the same PE Core L/14 backbone, training data, preprocessing, and comparable adaptation settings as CuRe? While Figure 4 supports continuous supervision over binary supervision, a matched comparison with prior detectors would help isolate the contribution of CuRe to the improvements in Tables 1 and 2.

  2. Performance across backbones. Appendix Table 8 reports 97.53% ROC-AUC for CuRe with PE Core L/14 but 80.13% with CLIP ViT-L/14 at LoRA rank 64. Could you extend the cross-backbone evaluation beyond AIGIBench and compare CuRe with strong methods such as GAPL under matched CLIP-based training conditions? This would help establish whether the advantage over previous methods persists independently of the pretrained representation.

4 comments
AuthorsOpenReview ↗

Backbone-matched comparisons with prior methods (PE Core and CLIP)

Thank you for the thoughtful questions and for the suggestions on additional experiments, which have made our comparisons more complete.

Q1 / W1: Comparison under the same PE Core L/14 backbone

Response: Thank you for the suggestion. We ported strong prior methods to PE Core L/14, using the same preprocessing and evaluation pipeline as CuRe, and evaluated them on all 10 benchmarks. For DRCT and DDA, the construction of the training data is part of the method, so we keep their official training data. The contributions of UnivFD and GAPL lie in the model and the training scheme, so we train them on the same data as CuRe.

Balanced accuracy (%); all methods use PE Core L/14.

MethodOriginal backboneTraining dataStandard avg.In-the-wild avg.Overall
UnivFD-PECLIP ViT-L/14GenImage SD v1.4 (same as CuRe)81.970.374.9
DRCT-PECLIP ViT-L/14SD v1.4 subset of DRCT-2M (official)88.273.979.6
DDA-PEDINOv2 ViT-L/14MSCOCO + DDA-aligned reconstructions (official)88.376.881.4
GAPL-PECLIP ViT-L/14GenImage SD v1.4 (same as CuRe)83.470.775.8
Binary LoRAPE Core L/14GenImage SD v1.4 (same as CuRe)82.069.074.2
CuRePE Core L/14GenImage SD v1.490.689.189.7

Overall, some methods benefit from the PE Core backbone, but CuRe still clearly outperforms all ported methods: its Overall accuracy is more than 8 points higher than the best of them, DDA-PE (89.7 vs. 81.4). In addition, the binary baseline in the table uses the same PE Core backbone, training data, and LoRA setting as CuRe, yet reaches only 74.2 Overall. The advantage of CuRe therefore cannot be attributed to the backbone alone.

Q2 / W2: Comparison with GAPL under the same CLIP backbone and training data

Response: The official GAPL weights are trained on different data from ours: GAPL is trained on Community Forensics, which covers about 4.7k generative models, whereas CuRe uses a single generator, GenImage SD v1.4. We therefore retrained GAPL with the same CLIP ViT-L/14 backbone and training data as CuRe, and compared the two on all 10 benchmarks.

MethodBackboneTraining dataStandard avg.In-the-wild avg.Overall
GAPL (retrained)CLIP ViT-L/14GenImage SD v1.481.870.274.8
CuReCLIP ViT-L/14GenImage SD v1.484.778.981.2
GAPL (official weights)CLIP ViT-L/14Prototype set + Community Forensics (about 4.7k generative models)95.074.482.6
CuRePE Core L/14GenImage SD v1.490.689.189.7

The backbone does have a clear effect on performance. However, under the same backbone and training data, CuRe still clearly outperforms GAPL (81.2 vs. 74.8).

We have added all the comparisons from both questions to the revised PDF (Appendix B.4, marked in blue). We hope this addresses your concern about the backbone, and we are happy to discuss further.

AuthorsOpenReview ↗

Follow-up on backbone-matched comparisons

Dear Reviewer 1XZ2,

Thank you again for raising the question of backbone fairness. In our response, we ported strong prior methods to PE Core L/14 and also retrained GAPL with the same CLIP ViT-L/14 backbone and training data as CuRe. CuRe stays clearly ahead in both cases (Appendix B.4). If you have any further questions, or if any part of our response is unclear, we would be glad to continue the discussion.

Best regards, The Authors

Reviewer_1XZ2OpenReview ↗

Official Comment

Thank you for the response. The backbone-matched results address most of my concerns.

Table 10 currently reports only the averages for the standard and in-the-wild benchmarks. I suggest reporting results for each of the four standard benchmarks and six in-the-wild benchmarks so readers can assess performance on individual datasets. I would also suggest the authors to merge Table 10 into Table 1 so readers can compare all methods and variants directly.

Overall, I find the idea of continuous regression of the mixing ratio interesting. I will raise my score to Weak Accept (3).

Dear Reviewer 1XZ2,

Thank you for reading our response and for raising your rating. We appreciate your suggestions on Table 10 and will revise the paper accordingly in the final version.

Thank you again for your time and constructive feedback.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.