acceptodds
Under review as a conference paper at ICLR 2027

Beyond 2D Artifacts: Geometric Cues for AI-Generated Image Detection

Abstract

AI-generated image detection remains challenging under unseen-generator shifts, in-the-wild settings, and repeated online dissemination. Existing image-domain cues can be generator-dependent and can degrade under resizing and recompression during online dissemination. We instead examine geometric coherence, motivated by the fact that real photographs arise from physical 3D scenes whereas AI-generated images need not satisfy the same physical constraints. To capture geometric coherence from a single image, we introduce GeoCue. GeoCue first considers point-normal consistency, which measures the agreement between normals implied by predicted 3D points and independently predicted surface normals. With a frozen geometry predictor, we observe that real photographs exhibit a characteristic range of point-normal consistency, whereas AI-generated images deviate from this range in generator-dependent ways. While point-normal consistency retains its discriminative signal under online dissemination, it can miss geometric changes that preserve point-normal agreement. GeoCue therefore further incorporates reference-relative depth and point deviations. GeoCue consistently outperforms prior methods across unseen-generator and in-the-wild benchmarks. Under repeated online dissemination involving resizing and JPEG/WebP re-encoding, GeoCue retains stable detection performance, while many existing detectors degrade substantially or exhibit unstable performance. These results demonstrate the robustness of GeoCue across generator shifts, in-the-wild data, and repeated online dissemination.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI-Generated Image Detection

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_kSqEOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces GeoCue, a detector that derives three image-level responses from monocular geometry predictions: disagreement between surface normals inferred from a predicted 3D point map and directly predicted normals, and the deviations of predicted depth and 3D points from those of a frozen reference geometry model. A trainable geometry predictor is optimized with margin-based objectives that encourage real-image responses to remain consistent with the reference while pushing AI-generated-image responses beyond specified margins. A linear classifier combines the three responses. The model is trained on 8,000 images and evaluated under cross-generator shifts, in-the-wild settings, and repeated resizing and JPEG/WebP re-encoding.

Critical strengths
  1. Interpretable and testable geometric responses. The proposed cues are explicitly defined, and the paper includes useful controls using the same MoGe-2 backbone, including a direct CLS-based classification alternative. This helps establish the empirical value of the geometric response formulation.

  2. Relevant robustness evaluation. The experiments model repeated online redistribution through several cumulative resizing and re-encoding sequences. Testing both detection accuracy and the stability of geometric responses under these transformations addresses an important practical limitation of methods based primarily on low-level image artifacts.

Critical weaknesses
  1. The interpretation of learned responses as intrinsic physical inconsistency remains insufficiently validated. The frozen-predictor observations in Figure 2 provide preliminary evidence that geometric predictions can differ between real and AI-generated images. However, the final training objectives explicitly encourage fake-image responses to diverge from the reference predictor. Therefore, larger deviations after training do not independently establish that the input image violates physical geometry; they may also reflect the learned discriminative objective, the geometry estimator's domain bias, or prediction errors. This distinction becomes especially relevant outside natural photography. Human-created illustrations can deliberately violate photographic perspective, while AI-generated images may depict coherent geometry. The authors should clarify whether their claims are restricted to photographic content and, if broader applicability is intended, evaluate style-matched human-created and AI-generated artwork. A non-geometric reference-versus-adapted-model control would also help establish how much of the gain is specifically attributable to geometry.

  2. Training-source confounding and substantial class-wise failures in some subsets. The paper uses only 8,000 training images, but these are drawn from ten comparatively recent generators, whereas the external baselines use different training sources and, often, older generator families (Table 7). This makes it difficult to attribute the cross-generator advantage specifically to the geometric method or to superior sample efficiency. Although the paper includes same-backbone internal comparisons, a stronger same-training-data detector baseline would improve the external comparison. Furthermore, the claim of balanced performance needs qualification: Table 10 reports only 10.3% real-image accuracy, versus 100% fake-image accuracy, on the StarGAN subset; Table 9 reports 68.8% real-image accuracy on Chameleon. These correspond to substantial false-positive rates at the paper's fixed 0.5 threshold. Strong mean AUC does not eliminate the need to characterize such operating-point failures.

Questions for authors
  1. Is GeoCue intended specifically for natural photographs, or also for illustrations, paintings, and other non-photographic content? If the latter, can the authors evaluate style-matched human-created and AI-generated artworks and report false-positive rates on the human-created subset? Can they compare against a non-geometric teacher–student or reference-deviation control to isolate the role of geometry?
  2. Can the authors retrain a strong conventional detector using the same 8,000 training images and a comparable model-selection protocol? Appendix A states that the training set does not overlap with the evaluation benchmarks; what checks were performed for exact and near-duplicate images or shared source content?
  3. How do the authors explain the very low real-image accuracy for StarGAN and the relatively high false-positive rate on Chameleon? The paper already uses a fixed 0.5 threshold, but could the authors additionally report threshold sensitivity and true-positive rates at low, validation-set-defined false-positive rates? Please also give the numerical values of the six loss weights in Eqs. (3)–(4) and explain precisely how invalid geometry pixels are excluded.

I would reconsider my assessment if these questions are addressed convincingly.

4 comments
AuthorsOpenReview ↗

Response to Reviewer kSqE

Dear Reviewer kSqE,

We sincerely thank you for recognizing the interpretability of GeoCue's geometric responses and the relevance of our evaluation under repeated online dissemination. We have conducted additional controlled experiments and analyses to address your concerns regarding the role of geometric cues, training-data fairness, and class-wise detection failures, while clarifying the study scope and implementation details.

 

1. Scope of the Study.

  • Clarification (Appendix E). GeoCue is designed to exploit geometric differences between natural photographs of real 3D scenes and AI-generated images. This scope is consistent with the physical image formation motivation presented in the Abstract and Introduction, as well as the evaluation settings of major benchmarks such as LSUN-Bedroom, AIGCDetectBench, Chameleon, and WildRF, which primarily use real photographs as the real-image class. Accordingly, distinguishing human-created paintings or illustrations from AI-generated artwork is beyond the scope of this study. We have clarified this scope in the revised manuscript.

 

2. Non-Geometric Reference-Deviation Control.

  • Experimental Setup. To examine whether GeoCue's performance gains can be explained solely by generic teacher–student feature differences, we introduced a non-geometric feature-deviation baseline. Using the same MoGe-2 backbone, LoRA configuration, and 8K training set, this baseline employs the cosine distance between encoder features from the frozen teacher and LoRA-adapted student as the detection signal, rather than deviations between geometric predictions.

  • Results (Appendix D, Table 16). GeoCue achieves a mean AUC of 88.3% across five benchmarks, outperforming the feature-deviation baseline (74.3%) by 14.0 percentage points. Its consistently higher AUC across all five benchmarks indicates that generic teacher–student feature differences alone cannot fully explain GeoCue's performance, supporting the effectiveness of explicit geometric responses as detection cues.

 

3. Controlled Comparison with Matched Training Data.

  • Experimental Setup. To control for differences in training data, we retrained UnivFD, C2P-CLIP, and AIDE on the same 8,000 AIGI-Holmes images used to train GeoCue. For a fair comparison under matched training data, each method was initialized from its pretrained backbone rather than its publicly released detector checkpoint, while following its original training procedure.

  • Training–Evaluation Overlap Check. We used separate datasets and generators for training and evaluation, and further examined potential image overlap beyond differences in dataset names. In particular, to prevent overlap between AIGI-Holmes and AIGCDetectBench, we cross-checked the real images in AIGCDetectBench against the original image IDs from COCO train2017 and val2017. Images sharing the same COCO source IDs were excluded from the training candidates, and the final 8K training set was constructed while maintaining balanced real/fake samples across generators. We also examined the data sources and composition of LSUN-Bedroom, Chameleon, LOKI, and WildRF to check the separation between training and evaluation data.

  • Results (Table 4). GeoCue achieves higher AUCs than the retrained baselines on LSUN-Bedroom (87.6 vs. 69.7), Chameleon (88.2 vs. 85.9), LOKI (81.9 vs. 70.7), and WildRF (94.3 vs. 90.0), demonstrating strong overall detection performance. Furthermore, its R.Acc./F.Acc. results on Chameleon (68.8/90.6), LOKI (65.7/77.6), WildRF (83.8/91.3), and AIGCDetectBench (77.3/89.3) demonstrate balanced performance across real and AI-generated images. These results show that GeoCue maintains strong generalization under matched training data, indicating that its gains cannot be explained solely by differences in training-set size or generator composition and supporting the effectiveness of geometric cues.

 

AuthorsOpenReview ↗

Response to Reviewer kSqE

4. Analysis of GAN Failure Cases.

  • Experimental Setup. To investigate the low real-image accuracy observed on StarGAN, we analyzed the differences in median responses between real and generated images for three geometric cues (cpnc_{pn}, rdepthr_{\mathrm{depth}}, and rpointr_{\mathrm{point}}). We used the original GeoCue model without additional training or threshold adjustment and compared these differences with those across the full AIGCDetectBench.

  • Results (Appendix E). On StarGAN, the response differences are approximately 0, 0.283, and 0.090, respectively, substantially smaller than 0.807, 1.274, and 0.457 across AIGCDetectBench. These findings are consistent with StarGAN's generation process, which modifies facial attributes such as hair color and age while largely preserving facial structure. This suggests that geometric cues may become less discriminative when generated images exhibit limited changes in geometric structure.

 

5. Threshold Sensitivity and False-Positive Reduction.

  • Experimental Setup. To examine the effect of threshold selection on false positives and detection performance, we calibrated a single threshold to a target FPR of 1% using a separate validation set, without retraining GeoCue. We then applied this threshold uniformly across three benchmarks and compared it with the default threshold of 0.5.

  • Results (Appendix D, Table 15). The calibrated threshold reduces FPR and narrows the gap between real-image and AI-generated-image accuracy across all three benchmarks. Notably, on LOKI, FPR decreases from 34.3% to 27.3%, while the class-wise accuracy gap narrows from 11.9 to 2.8 percentage points. Although this adjustment also leads to some reduction in TPR, these results demonstrate that validation-based threshold calibration can reduce false positives and improve class-wise balance without additional training.

 

6. Loss Weights and Valid-Pixel Masking.

  • Loss Weights (Sec. 3.4). The six loss weights in Eqs. (3)–(4) are set to (λdreal,λpreal,λpnreal)=(1.0,0.5,0.3)(\lambda_d^{\mathrm{real}}, \lambda_p^{\mathrm{real}}, \lambda_{pn}^{\mathrm{real}})=(1.0, 0.5, 0.3) for real images and (λdAIGI,λpAIGI,λpnAIGI)=(1.0,1.0,1.0)(\lambda_d^{\mathrm{AIGI}}, \lambda_p^{\mathrm{AIGI}}, \lambda_{pn}^{\mathrm{AIGI}})=(1.0, 1.0, 1.0) for AI-generated images.

  • Valid-Pixel Masking. We retain only pixels where the frozen reference predictor's mask confidence exceeds 0.5 and the depth, normal, and 3D-point predictions from both the reference and cue predictors are finite (neither NaN nor Inf). Invalid pixels are excluded from geometric response computation, and the responses are averaged over valid pixels only. For scale normalization, we use the median of positive reference depth values among valid pixels.

 

We hope these clarifications address your concerns, and we are glad to run any further tests you would find useful. If your questions are addressed convincingly, we would be grateful if you would consider it in your final assessment.

Best, The Authors

Thank you for your response; after careful consideration, I have decided to maintain my original rating.

AuthorsOpenReview ↗

Response to Reviewer kSqE

Dear Reviewer kSqE,

We sincerely thank you for recognizing the interpretability of GeoCue's geometric responses and the practical relevance of our evaluation under repeated online dissemination. We also appreciate your detailed and constructive feedback.

In our previous response, we addressed your concerns through additional experiments and clarifications, including a non-geometric teacher–student comparison, comparisons with existing detectors trained on matched data, an analysis of the StarGAN failure case, threshold calibration, and implementation details.

 

1. Scope of Applicability.

GeoCue is designed to distinguish AI-generated images from natural photographs of real 3D scenes. Accordingly, distinguishing human-created artwork from AI-generated artwork in non-photographic content, such as illustrations and paintings, is beyond the scope of this study. We have explicitly clarified this scope in Appendix E (Scope of Applicability). To further communicate the intended scope of our work, we propose revising the paper title as follows: Beyond 2D Artifacts: Geometric Cues for Distinguishing Natural Photographs from AI-Generated Images

2. Limitations on Attribute-Level Manipulation.

Separately, even within natural-photograph-based generation and editing, geometric cues alone may provide limited discriminative power when certain attributes are modified while the overall facial geometry is largely preserved, as in StarGAN, which modifies attributes such as hair color and age.

In this regard, Wang et al. (2024), in their DEEPFAKER study, evaluated multiple existing detectors on various attribute-editing methods, including StarGAN, AttGAN, and HiSD. Their results showed that detection performance varied across manipulation types even for the same detector, with several detectors exhibiting limited generalization across these editing methods.

Furthermore, Sarkar et al. (CVPR 2024) exploit scene-level geometric cues, including perspective structure, straight-line geometry, and object–shadow relationships, to detect geometric inconsistencies in AI-generated images. When such geometric relationships are preserved under localized attribute modifications, these cues alone may provide limited discriminative evidence.

Furthermore, Zhang et al. (2026) explored complementary cues such as facial dynamics, beyond static visual features, for fine-grained facial manipulation detection in videos. Together, these studies suggest that the effectiveness of detection cues can vary depending on the generative models and manipulation types.

3. Effectiveness and Robustness of GeoCue.

Despite these limitations, GeoCue demonstrates competitive performance across diverse generative models and real-world settings within the intended scope of natural-photograph-based detection.

Specifically, GeoCue achieves a mean AUC of 89.3% across 16 generators in AIGCDetectBench and 87.6% across 10 generators in LSUN-Bedroom. On the in-the-wild benchmarks, GeoCue achieves AUCs of 88.2%, 81.9%, and 94.3% on Chameleon, LOKI, and WildRF, respectively, yielding the highest mean AUC of 88.1% across the three benchmarks.

Moreover, under three online dissemination settings involving repeated resizing and JPEG/WebP re-encoding, GeoCue maintains more stable performance than many existing detectors. These results support the ability of GeoCue's compact and interpretable geometric cues to generalize across diverse datasets and provide relative robustness under practical online sharing conditions.

4. Future Directions.

An important direction for future work is to complement geometric cues with attribute-aware visual cues that capture localized appearance changes arising from fine-grained attribute manipulations, particularly when the underlying geometric structure remains unchanged. Integrating these complementary cues with GeoCue's geometric cues may help improve its generalization to diverse generative models and fine-grained manipulation types.

 

We hope that these additional clarifications and revisions help address your concerns. If there are any aspects that remain insufficiently addressed or require further validation, we would greatly appreciate it if you could let us know specifically. We welcome any further comments or suggestions and would be glad to provide additional clarifications or analyses to further strengthen our work.

 

Thank you again for your careful evaluation and constructive feedback.

Best,
The Authors

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

The paper proposes GeoCue, a detector for AI-generated images based on predicted 3D geometry. The motivation is that real photos come from physical 3D scenes, while generated images do not have to obey the same physical rules. In GeoCue, two copies of the MoGe-2 geometry model are used, which are a frozen reference copy and a cue copy that is fine-tuned with LoRA which computes three scores per image, which are a point-normal consistency cpnc_{pn}, how well the surface normals implied by the predicted 3D points agree with the directly predicted normals and the depth deviation rdepthr_{depth} and point deviation rpointr_{point} which shows how far the cue copy’s depth and 3D points move away from the frozen reference because the cue copy is trained with a margin loss the real images are pushed toward good consistency and small deviation and AI-generated images are pushed beyond fixed margins, this shows that the three scores are then fed to a linear classifier and training uses 8,000 images (10 generators) from AIGI-Holmes. The results from "In the wild" show the highest mean AUC on Chameleon, LOKI and WildRF (88.1 vs. 83.3 for the best baseline, DEAR-c), with more balanced accuracy on real and fake images than most baselines and unseen generators which show the highest mean AUC on AIGCDetectBench (89.3) and LSUN-Bedroom (87.6) and under three sequences of repeated resizing and JPEG/WebP re-encoding, GeoCue’s AUC drops less than most baselines.

Critical strengths
  • GeoCue reaches 88.2 AUC on Chameleon, compared with 74.4 for the best baseline (AIDE), and 94.3 on WildRF. It does this with only 8K training images. It also avoids the extreme class bias many baselines show. For example, on Chameleon, CNNSpot labels almost every image as real (R.Acc 99.6, F.Acc 0.0), while GeoCue has R.Acc 68.8 and F.Acc 90.6 (Table 9).

  • The three dissemination sequences are fully specified (Table 8), and GeoCue is compared against up to 13 detectors across rounds (Figs. 6 and 9). GeoCue degrades less than most baselines. The idea of using a monocular geometry model’s outputs as a compact, three-number detection signal is new and well illustrated (Figs. 3 and 7).

Critical weaknesses

The paper does not show that the gains come from geometry rather than from fine-tuning a strong backbone. because paper’s central claim is that geometric coherence, grounded in physical image formation, is the source of detection evidence. Several results in the paper point the other way, which i will outline below for the authors:

1. The cues are trained outputs, not measured properties as the cue predictor is trained with real/fake labels so that fake images get large cpnc_{pn} and large deviation from the frozen copy (Eq. 3-5; LoRA on the last 12 blocks, App. A). So rdepthr_{depth} and rpointr_{point} are not geometric properties of the image itself. They are outputs the network is trained to move for fake images. The abstract and introduction present the cues as observations (“we observe…”) and do not mention this training; it first appears at the end of (Section 3.4). The untrained signal is weak as shown with the frozen predictor, the point-normal differences in Fig. 2a lie between about -0.04 and +0.06, and the sign varies by generator. The trained margin is mpn=0.5m_{pn} = 0.5. Frozen MoGe-2 reaches only 63.6 AUC on LOKI (Table 5b).

2. Reversing the physical hypothesis still beats every baseline as shown in Table 5a, training fake images toward better point-normal agreement than real images still gives 85.8 AUC on Chameleon. Reversing every term gives 80.6. Both exceed every baseline on Chameleon in Table 1 (best: AIDE, 74.4). If the opposite of the physical hypothesis also produces state-of-the-art results, the gain likely comes from fine-tuning MoGe-2 on this training set, not from physical image formation. In Figs. 7 and 10, fake images show low consistency and high depth deviation over almost the whole image, including backgrounds. This fits a global “fake” signal better than localized geometric errors. The paper’s own control, MoGe-2 with CLS fine-tuning (Table 5b), trails GeoCue by only 3.4 AUC on LOKI (78.5 vs. 81.9) and 2.1 on WildRF (92.2 vs. 94.3). This comes from a single run, it is not reported on Chameleon or under dissemination, and its training setup is not described. I don’t think this gap is enough to support the claim that explicit geometric cues are the key ingredient.

3. The baseline comparisons are not controlled, and the headline claims go beyond the results as show by the training data, In Table 7 lists the baselines’ training data and frames the difference as training scale. The larger difference is how recent the generators are. GeoCue trains on modern generators (FLUX, SD3.5, PixArt, Janus and others), while most baselines are trained on ProGAN/LSUN or LDM. The in-the-wild benchmarks contain modern generators. No baseline is retrained on the same 8K images, apart from the MoGe-2 CLS model, so the effect of the method cannot be separated from the effect of the training data. Baseline selection changes between tables as the DEAR-c, the strongest baseline in Table 1 (83.3), is missing from Fig. 4, Table 9 and Fig. 6. Table 2 shows 6 baselines and Table 10 shows 9. NPR, AIDE, FerretNet and DFFreq appear in neither. In the AIGCDetectBench panels of Figs. 6 and 9, one baseline curve, which appears from the legend to be AIDE, starts at about 95–98 AUC at Round 0. That is above GeoCue’s 89.3, yet AIDE is not reported in Table 2 or Table 10. The wording that the “Consistently outperforms” is not accurate. On LOKI, GeoCue (81.9) ranks fifth, behind DEAR-c (85.9), DEAR-r (85.8), AIDE (82.6) and DFFreq (82.6). On LSUN, it leads C2P-CLIP by only 0.7 AUC and has lower mean AP (84.6 vs. 86.8). “Retains stable detection performance” is not accurate as shown in Fig. 9, GeoCue falls from about 88-89 to about 71-77 AUC by Round 5. That is less degradation than many baselines, but it is not stable. App. D’s wording, “relatively stable,” is accurate, and the abstract and introduction should match it.

There are some minor details that the authors should fix as well which i will outline as below:

  • The abstract and intro do not say the cue predictor is trained with real/fake labels.
  • R.Acc. and F.Acc. are used from L087 but only defined in App. A.
  • “Low cross-benchmark variation” (L088-089) has no supporting number.
  • (Section 3.1) (L205-212) uses stability under re-compression to motivate a deviation between two predictors, which is a different quantity.
  • The symbols c_dn and r_normal in Table 4a are undefined.
  • Names are inconsistent (WhichFaceIsReal/WFIR, CNN-Spot/CNNSpot).
Questions for authors

1. In Table 5a, reversing the point-normal objective still gives 85.8 AUC on Chameleon, and reversing everything gives 80.6. Both are above every baseline in Table 1. How do you reconcile this with the claim that the physical direction of the cues is what makes GeoCue work?

2. Can you retrain two or three strong baselines (for example UnivFD, C2P-CLIP or AIDE) on the same 8K AIGI-Holmes images, or train GeoCue under the standard ProGAN/LSUN protocol? Were the baselines evaluated with their released checkpoints?

3. How was the MoGe-2 CLS fine-tuning baseline trained (LoRA or full fine-tuning, which layers, same data and epochs)? What is “MoGe-2 (Frozen)” in Table 5b: frozen geometric cues, or a linear probe on frozen features? Please also report the CLS baseline on Chameleon and under the dissemination sequences.

4. What is AIDE’s AUC on the original AIGCDetectBench images, and why is it not in Tables 2 and 10? Why is DEAR-c left out of Fig. 4, Table 9 and Fig. 6?

5. What are the λ values in Eq. 3-4, and how was m_d = m_p = 0.1875 chosen? Were the margins or the cue set tuned using any evaluation benchmark? Tables 4a and 4b use LOKI and WildRF, and Fig. 2 uses AIGCDetectBench generators.

6. Fig. 2b measures deviation with a frozen predictor across dissemination rounds. What is this deviation relative to, and how does it relate to rdepthr_{depth}, which compares two predictors on the same image?

7. Please report variance over several training seeds, given the small gaps on LSUN (+0.7 AUC) and AIGCDetectBench (+1.8 AUC).

5 comments
AuthorsOpenReview ↗

Response to Reviewer HdEG

Dear Reviewer HdEG,

We sincerely thank you for your detailed review and for recognizing GeoCue's strong in-the-wild performance and the novelty of its compact geometric-cue formulation. We have conducted additional experiments and analyses to address your concerns regarding the role of geometric cues, baseline comparisons, and training-seed variability.

 

1. Objective Direction (Table 6(a)).

  • Motivation and Objective (Fig. 2). As shown in Fig. 2(a), real images exhibit a relatively consistent range of point-normal consistency responses, whereas AI-generated images deviate from this range in generator-dependent directions. Fig. 2(b) further shows that geometric predictions for real images remain relatively stable under repeated online dissemination, compared to those for AI-generated images. Motivated by these observations, GeoCue defines its training objective based on the observed geometric responses of real images rather than the direction of deviations in AI-generated images. Specifically, we encourage stronger point-normal agreement and smaller reference-relative deviations for real images, while pushing AI-generated responses beyond the corresponding margins.

  • Reversed Objectives. As illustrated in Fig. 2(a), AI-generated responses deviate from the real-image range in generator-dependent directions. This suggests that the two classes can be distinguished without enforcing a specific direction of deviation. Thus, reversing the objective can still produce discriminative responses. However, GeoCue, designed around the relative stability of real-image geometric predictions observed in Fig. 2(b), achieves higher AUC than both reversed configurations in Table 6a: 88.2 versus 85.8 (Reversed Consistency) and 80.6 (Reversed All) on Chameleon, and 81.9 versus 78.3 and 70.0 on LOKI, respectively. These results indicate that while reversed objectives can also learn discriminative signals, aligning the objective with the observed geometric response characteristics of real images is more effective.

 

2. Controlled Comparison with Matched Training Data.

  • Original Evaluation. The original comparisons used the publicly released detector checkpoints.

  • Matched Training Setup. We retrained UnivFD, C2P-CLIP, and AIDE on the same 8,000 AIGI-Holmes images used for GeoCue. To avoid confounding effects from the released detectors' training data, each method was initialized from its pretrained backbone rather than its detector checkpoint, following its original training procedure.

  • Results (Table 4). GeoCue achieves higher AUCs on LSUN-Bedroom (87.6 vs. 69.7), Chameleon (88.2 vs. 85.9), LOKI (81.9 vs. 70.7), and WildRF (94.3 vs. 90.0), with comparable performance on AIGCDetectBench (89.3 vs. 89.6). These results show that GeoCue maintains its advantage on most benchmarks under matched training data and generator composition, indicating that its gains cannot be explained solely by differences in training data.

 

AuthorsOpenReview ↗

Response to Reviewer HdEG

3. MoGe-2 Baseline Clarification.

  • Baseline Settings. MoGe-2 (CLS FT) applies rank-16 LoRA to the last 12 transformer blocks and jointly trains a CLS-token-based linear classifier, using the same 8K AIGI-Holmes training set and number of training epochs as GeoCue. MoGe-2 (Frozen) keeps the backbone frozen and trains only a linear classifier on CLS features.

  • Chameleon Comparison. On the original Chameleon images, MoGe-2 (Frozen), CLS FT, and GeoCue achieve AUCs of 69.40, 89.49, and 88.18, respectively. After five dissemination rounds, their AUCs are 58.69/75.20/77.12. Thus, although CLS FT performs slightly better on the original images, GeoCue outperforms CLS FT after dissemination, with a smaller performance drop.

  • Additional Results (Table 6(b), Table 16). GeoCue also outperforms CLS FT on LOKI (81.9 vs. 78.5) and WildRF (94.3 vs. 92.2), as well as a non-geometric teacher–student feature-deviation baseline across all five benchmarks. These results support the contribution of explicit geometric responses beyond pretrained visual representations, such as those used in CLIP-based detectors, to cross-dataset generalization and robustness under repeated dissemination.

 

4. Additional Baseline Results.

  • AIDE (Fig. 6, Fig. 8). On original AIGCDetectBench images, AIDE achieves an AUC of 98.6%, exceeding GeoCue's 89.3%, as shown at Round 0 in Fig. 6. However, AIDE exhibits a substantially larger performance decline under repeated online dissemination, whereas GeoCue shows a relatively smaller decline. These comparisons are presented in Fig. 6 and Appendix C (Fig. 8).

  • DEAR-c (Table 11, Fig. 8). Figs. 4 and 6 visualize representative baselines, while DEAR-c's detailed results are reported in Appendix B (Table 11) and its dissemination results in Appendix C (Fig. 8). DEAR-c achieves AUC/R.Acc./F.Acc. of 72.1/76.2/59.2 on Chameleon, 85.9/89.3/74.5 on LOKI, and 92.0/96.9/77.1 on WildRF. Across these benchmarks, GeoCue achieves a higher mean AUC (88.1% vs. 83.3%).

 

5. Loss Weights, Margin Selection, and Geometric Cues.

  • Loss Weights (Sec. 3.4). The six loss weights in Eqs. (3)–(4) are (λdreal,λpreal,λpnreal)=(1.0,0.5,0.3)(\lambda_d^{\mathrm{real}}, \lambda_p^{\mathrm{real}}, \lambda_{pn}^{\mathrm{real}})=(1.0,0.5,0.3) for real images and (λdAIGI,λpAIGI,λpnAIGI)=(1.0,1.0,1.0)(\lambda_d^{\mathrm{AIGI}}, \lambda_p^{\mathrm{AIGI}}, \lambda_{pn}^{\mathrm{AIGI}})=(1.0,1.0,1.0) for AI-generated images.

  • Margin Selection (Table 5(b)). We set mpn=0.5m_{pn}=0.5 based on point-normal consistency responses of real training images measured with the frozen geometry predictor, with 99.65% falling below this margin. For depth and point deviations, we use md=mp=0.1875m_d=m_p=0.1875. Table 5(b) shows that the default 1.0×1.0\times margin scale achieves the highest mean AUC across LOKI and WildRF.

  • Geometric Cue Selection (Table 5(a)). The three geometric cues are designed based on point-normal consistency and reference-relative deviation, without selecting the cue set using evaluation benchmarks. Table 5(a) evaluates their complementarity.

 

6. Clarification of Fig. 2(b).

  • Measurement in Fig. 2(b). Fig. 2(b) measures the mean absolute difference between depth predictions from the same frozen MoGe-2 predictor applied to the original image (Round 0) and its cumulatively disseminated version at each round.

  • Relation to Eq. (2). In contrast, rdepthr_{\mathrm{depth}} compares predictions from the frozen reference and LoRA-adapted cue predictors on the same image. Thus, Fig. 2(b) illustrates the relative stability of real-image geometric predictions under dissemination as design motivation, while Eq. (2) defines the learned geometric detection cue motivated by this observation.

 

7. Robustness Across Training Seeds.

  • Experimental Setup. We independently trained GeoCue with four random seeds and evaluated AUC on LSUN and AIGCDetectBench. For AIGCDetectBench, we averaged AUC across all 16 generators and report the mean and sample standard deviation over the four seeds.

  • Results (Appendix E). GeoCue achieves 87.57±0.9287.57 \pm 0.92% on LSUN and 90.25±1.2390.25 \pm 1.23% on AIGCDetectBench. Seed-dependent variation differs across generators, with larger fluctuations for some GAN models. StarGAN shows the largest variation (69.84±12.3069.84 \pm 12.30) and, as discussed in Appendix E, substantially smaller real–fake geometric response differences than the overall AIGCDetectBench. This suggests greater seed sensitivity when geometric cues have limited discriminability.

 

We hope these responses address your concerns and welcome any further questions or feedback.

Best, The Authors

Thank you for the revision. The non-geometric teacher–student control (Table 16) and the matched-data comparison (Table 4) substantially strengthen the paper, and the loss weights and metric definitions are now clear and answers most of my concerns.

But i would like the authors to clarify the remaining points:

(1) The seed results described in the response (LSUN 87.57 ± 0.92, AIGCDetectBench 90.25 ± 1.23) do not appear in the revised PDF; Appendix E contains only the StarGAN limitation. Please add them, ideally also for the in-the-wild benchmarks and for Table 6.

(2) The Chameleon comparison with CLS FT reports only the post-dissemination result (Sequence C). Please also report the original-image result (89.49 vs. 88.18, as given in the response) and the other two sequences.

(3) Table 15 shows a real-image FPR of 52.9% on LSUN and 34.3% on LOKI, still 48.8% and 27.3% after calibrating to 1% validation FPR. Please discuss this in the main text, since it affects the class-balance claim.

(4) The Fig. 2(b) clarification from the response is not in Section 3.1. Please state there that Fig. 2(b) measures frozen-predictor change across dissemination rounds, which differs from r_depth in Eq. (2).

(5) Please state in the abstract and introduction that the cue predictor is trained with real/fake labels, and revise “retains stable detection performance” in the abstract (AUC drops by roughly 10–17 points by Round 5 in Fig. 8).

(6) Minor: please add AIDE to Tables 2 and 12 and DEAR-c to Fig. 4; explain the choice of m_d = m_p = 0.1875; fix the column order in Table 7 and state whether its baselines are released or retrained; cite Ricker et al. (Towards the Detection of Diffusion Model Deepfakes) for the LSUN-Bedroom generated set; and support or remove “low cross-benchmark variation” (L088–089).

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

The paper proposes GeoCue, which uses single-image geometry predictions to build detection features that distinguish real from AI-generated images. Trained on 8,000 images from 10 generators (AIGI-Holmes) and evaluated on AIGCDetectBench, LSUN-Bedroom, Chameleon, LOKI, and WildRF, it achieves strong average AUC. The key idea is leveraging consistency across geometry predictions and pre/post-adaptation deviations for transferable detection, but the link between gains and true physical geometric errors needs more direct validation.

Critical strengths
  1. The method shows good cross-dataset detection performance, evaluated under various image degradations.
  2. The detection approach is simple, and controlled experiments suggest the proposed method drives the observed gains.
Critical weaknesses
  1. The comparative experiments may be unfair because methods use different training data. Table 7 shows GeoCue trained on data from 10 generators (e.g., FLUX, Janus, SD35-L), while many baselines use ProGAN/LDM-style data, so training sets are not matched. Common benchmarks such as GenImage and DRCT-2M are also missing. Fewer training images does not necessarily mean easier or harder training; generator diversity also matters.

  2. Averages hide failure cases. In Table 10, GeoCue scores only 59.0 AUC on StarGAN with R.Acc./F.Acc. = 10.3/100.0, implying ~89.7% of real images are misclassified as fake. R.Acc. is also low on ProGAN/StyleGAN/StyleGAN2 (36.2/47.7/49.4), indicating high false-positive rates. More analysis of these results is needed.

  3. Modern generators can be guided by edge/depth/normal maps. If a model uses normal maps from real images for guidance, would GeoCue fail or degrade? Please report experiments on this setting.

  4. What is the computational overhead? Since the method computes geometric quantities, please report runtime and per-step timing for the geometric computations.

Questions for authors

See weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer vBmL

Dear Reviewer vBmL,

We sincerely thank the reviewer for recognizing GeoCue's strong cross-dataset performance and robustness under image degradations. We address the concerns through additional experiments and analyses on training-data fairness, GAN failure cases, geometry-guided generation, and computational overhead.

 

1. Comparison with Matched Training Data.

  • Matched Training Setup. We retrained UnivFD, C2P-CLIP, and AIDE on the same 8,000 images used for GeoCue, initializing each method from its pretrained backbone and following its original training procedure.

  • AUC Comparison (Table 4). GeoCue outperforms the retrained baselines on LSUN-Bedroom (87.6 vs. 69.7), Chameleon (88.2 vs. 85.9), LOKI (81.9 vs. 70.7), and WildRF (94.3 vs. 90.0).

  • Class-Wise Performance (Table 4). GeoCue maintains balanced R.Acc./F.Acc. on Chameleon (68.8/90.6), LOKI (65.7/77.6), WildRF (83.8/91.3), and AIGCDetectBench (77.3/89.3). These results show that GeoCue maintains strong generalization under matched training data, indicating that its gains cannot be explained solely by differences in training-set size or generator composition and supporting the effectiveness of geometric cues.

  • Benchmark Coverage. We evaluate five cross-generator and in-the-wild benchmarks. All eight GenImage generators overlap with AIGCDetectBench, while DRCT-2M shares diffusion generators including SD1.4, SD1.5, and LDM. Therefore, to minimize redundancy in generator coverage while evaluating diverse generation methods and real-world conditions, we did not separately include these two benchmarks.

 

2. Analysis of GAN Failure Cases.

  • Geometric Response Analysis (Appendix E). Using the original GeoCue model without retraining or threshold adjustment, we compared real–fake median response differences for three cues (cpnc_{pn}, rdepthr_{\mathrm{depth}}, and rpointr_{\mathrm{point}}). On StarGAN, these differences are approximately 0, 0.283, and 0.090, substantially smaller than 0.807, 1.274, and 0.457 across AIGCDetectBench.

  • Interpretation. StarGAN modifies facial attributes while largely preserving facial structure. The smaller geometric response differences suggest that GeoCue's discriminability may decrease when generated images exhibit limited geometric changes.

 

3. Geometry-Guided Generation.

  • Experimental Setup. Using 500 COCO real images and matched captions, we compared SD1.5 text-only generation with ControlNet depth-guided (MiDaS) and normal-guided (NormalBae) generation. GeoCue was evaluated without retraining or threshold adjustment, and baselines used their official checkpoints.

  • Geometric Response Analysis (Table 7). Depth and normal guidance achieved a Spearman correlation of 0.878 and cosine similarity of 0.913, respectively. Despite this alignment, rdepthr_{\mathrm{depth}} increased from 1.2745 to 1.4153/1.4483, and rpointr_{\mathrm{point}} from 0.4866 to 0.5209/0.5240. This suggests that although ControlNet incorporates the input depth/normal structure, it does not directly constrain the differences between the two geometry predictors evaluated on the same image.

  • Detection Performance (Table 7). GeoCue's AUC increased from 96.50% (text-only) to 97.71% (depth-guided) and 98.19% (normal-guided). F.Acc. also improved from 78.60% to 82.20% and 82.80%, respectively. GeoCue thus retains its geometric discriminability without performance degradation under the evaluated SD1.5/ControlNet settings.

 

4. Computational Overhead.

  • Experimental Setup. We measured end-to-end and stage-wise latency on 1,000 images (100 real and 100 AI-generated per benchmark across five benchmarks), using the final checkpoint on a single NVIDIA RTX PRO 6000 Blackwell GPU at 448 × 448 resolution, excluding preprocessing.

  • Latency Breakdown (Appendix D, Table 17). GeoCue achieves 59.742 ms/image end-to-end latency, with two geometry predictor forwards accounting for 58.625 ms (98.1%). The three geometric cues require only 0.362 ms (0.6%), valid-mask and scale normalization 0.539 ms (0.9%), and pooling/classification 0.217 ms (0.4%).

  • Additional Overhead. Total processing cost excluding predictor forwards is 1.118 ms/image (1.9%), indicating that inference cost is dominated by geometry prediction rather than geometric cue computation.

 

We hope these clarifications address all your concerns, and we are glad to clarify anything further.

Best, The Authors

AuthorsOpenReview ↗

Follow-up to Reviewer vBmL

Dear Reviewer vBmL,

We sincerely thank Reviewer vBmL for the thoughtful and constructive feedback, which has helped us improve the clarity and quality of our work.

We hope our responses adequately address your concerns, and we would welcome any further feedback during the discussion period. If any aspects require further clarification, please let us know so that we can address them.

Best, The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their thoughtful and constructive feedback. We appreciate their recognition of GeoCue's strong cross-dataset and in-the-wild detection performance (vBmL, HdEG), its simple and interpretable geometric-cue formulation (vBmL, HdEG, kSqE), and the practical relevance of its robustness evaluation under repeated online dissemination (vBmL, HdEG, kSqE). Below, we summarize the main contribution and revisions made in response to their comments. All section, table, and figure references refer to the revised PDF, with changes highlighted in blue.

Matched-Data Baseline Comparisons (vBmL, HdEG, kSqE). We retrained UnivFD, C2P-CLIP, and AIDE on the same 8,000-image training set as GeoCue and clarified baseline training protocols (Table 4; Appendix A). GeoCue achieves higher AUC on diverse benchmarks, supporting the effectiveness of compact geometric cues for cross-dataset detection under matched training conditions.

Geometry-Guided Generation (vBmL). We evaluated GeoCue under real-image depth- and normal-guided generation using SD1.5 and ControlNet (Table 7). Although the generated images follow the provided geometric guidance, this does not ensure consistency across the geometric predictions evaluated by GeoCue, leaving reference-relative depth and point deviations that remain discriminative.

Geometric-Cue Validation (HdEG, kSqE). Given the effectiveness of pretrained visual features in detectors such as UnivFD, we examined whether CLS-feature classification or generic teacher–student feature deviations could explain GeoCue's gains (Table 6(b); Appendix D, Table 16). The results show that explicit geometric responses provide valuable detection evidence beyond pretrained feature-based representations, contributing to stronger cross-dataset generalization and robustness under repeated online dissemination.

Inference Cost (vBmL). We reported end-to-end and stage-wise inference latency (Appendix D, Table 17). The results show that geometry prediction dominates inference cost, while computing the three geometric cues requires only 0.362 ms/image (0.6%).

Failure Analysis and Limitations (vBmL, HdEG, kSqE). We analyzed GeoCue's weaker performance on StarGAN, which modifies facial attributes such as hair color and age while largely preserving facial structure (Appendix E). The substantially smaller real–fake geometric-cue differences are consistent with our hypothesis that geometric information provides discriminative evidence, while also revealing GeoCue's limitations when generated images exhibit limited geometric changes. We also evaluated validation-based threshold calibration to reduce false positives (Appendix D, Table 15), clarified the study's scope and limitations (Appendix E), and reported training-seed variability (Appendix E).

 

We hope these additional results and revisions address the reviewers' concerns.

Best, 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.