acceptodds
Withdrawn from ICLR 2027

Generalizable AI-Generated Image Detection via Neighborhood Dependency Representation and Hierarchical Dependency Modeling

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing AI-generated image detectors often degrade on unseen generators because synthesis traces exhibit substantial variations across generation paradigms. Most existing methods focus on generator-specific artifacts or model forensic evidence from a single perspective, limiting their ability to capture transferable synthesis traces. We observe that although artifact appearances differ, different generators consistently alter image dependency structures across spatial and spectral scales. Based on this observation, we propose a hierarchical dependency modeling framework for generalizable AI-generated image detection. Specifically, we introduce the Neighborhood Dependency Representation (NDR), which reconstructs each pixel from its local context and extracts residual dependency variations to reduce semantic bias and reveal synthesis-related irregularities. Building upon NDR, the local subband dependency branch and global spectral dependency branch capture complementary local and global dependency structures, respectively. A bidirectional confidence guided fusion module further adaptively integrates these representations for robust classification. Extensive experiments on AIGCDetectBenchmark, ForenSynths, and WildRF demonstrate superior cross-generator and real-world generalization, showing the effectiveness of dependency-based forensic representation learning.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI-generated image detection, Deepfake detection, Image forensics, Generalization, Frequency analysis, Dependency representation

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_ZCX2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an AI-generated image detection framework that first extracts neighborhood median residuals (NDR), then processes them through local wavelet and global Fourier branches, and performs classification using bidirectional adaptive fusion. Trained using ProGAN as the sole source of generated images, the model outperforms the compared methods on ForenSynths, AIGCDetectBenchmark, and WildRF.

Critical strengths

1.The framework combines local and global forensic cues, and the single-generator training setup is appropriate for evaluating cross-generator generalization. 2.The method achieves the best average results among the compared methods on three benchmarks, including the real-world social-media dataset WildRF.

Critical weaknesses

1.Limited conceptual novelty over NPR. NPR already exploits neighboring-pixel relationships for generated image detection, so the dependency-oriented perspective emphasized in this paper substantially overlaps with prior work. The clearer differences are the median-based residual operator and the subsequent dual-branch architecture. 2.The contribution of NDR itself is not sufficiently isolated. Removing NDR only demonstrates that residual preprocessing is useful. Without comparisons replacing NDR with NPR or standard filtering residuals under the same architecture, the specific advantage of the proposed representation remains unclear. 3.The robustness evidence remains limited. Accuracy on WildRF is 67.1%, only 1.2 percentage points above the runner-up. The stronger claims about robust real-world generalization therefore require additional support.

Questions for authors

1.With the downstream architecture and training setup fixed, how does NDR compare with NPR, standard median-filter residuals, and mean-filter residuals? 2.What evidence distinguishes the proposed “dependency irregularities” from the local residual artifacts exploited by NPR and related methods?

1 comment
AuthorsOpenReview ↗

Response to Reviewer ZCX2

Thank you for your constructive feedback and for recognizing the complementary local/global design and the single-generator training protocol. We respond below. The additional experiments are included in the revised manuscript's appendix.

Q1. Controlled comparison with NPR and standard filtering residuals

We agree that removing NDR alone does not establish its advantage over alternative residual representations. We therefore fixed the downstream architecture and training settings and replaced only the front-end representation (Appendix C.1, Table 8). Entries below are mAcc./mAP (%).

Front-end representationAIGCDetectBenchmarkForenSynthsWildRF
Standard mean-filter residual82.3 / 91.678.3 / 89.260.4 / 67.6
Standard median-filter residual86.1 / 95.185.0 / 94.664.2 / 67.5
NPR86.1 / 93.574.2 / 81.258.5 / 66.8
NDR (ours)92.2 / 96.984.8 / 93.567.1 / 71.0

NDR achieves the best mAcc and mAP on AIGCDetectBenchmark and WildRF. Compared with the standard median-filter residual, the gains are 6.1/1.8 and 2.9/3.5 percentage points, respectively, while performance remains close to the best on ForenSynths.

The key to NDR is highlighting dependencies between the center pixel and its neighbors. Unlike standard median filtering, we set the center pixel to zero before median aggregation, making the reconstruction independent of its original value and more dependent on surrounding pixels. The difference between the original pixel and this reconstruction explicitly measures its deviation from neighborhood context, highlighting potential synthesis-induced dependency irregularities. These explanations and comparisons are included in Appendix C.1.

Q2. Distinction from NPR and the meaning of “dependency irregularities”

NPR primarily targets local pixel correlations associated with upsampling operations, using differences within local grids relative to a reference pixel at a fixed sampling position. NDR instead constructs a contextual baseline through neighborhood median reconstruction. Here, “dependency irregularities” has a concrete representational meaning: deviations between original pixel values and their neighborhood-supported reconstruction baselines.

Building on NDR, we model these irregularities at complementary local and global scales. The local subband dependency branch (LSDB) captures directional texture inconsistencies and cross-subband associations through wavelet decomposition and local-window interactions. The global spectral dependency branch (GSDB) captures image-wide spectral correlations, periodic responses, and long-range contextual information through Fourier-domain modeling. Bidirectional confidence-guided fusion (BCGF) then mutually calibrates and adaptively integrates these features according to the input image.

The controlled comparison in Table 8 shows that NDR outperforms NPR in both mAcc and mAP on all three benchmarks under the same downstream architecture and training settings, supporting the benefit of neighborhood-context reconstruction.

Furthermore, with the same NDR input and without BCGF, LSDB alone achieves 86.1% mAcc on AIGCDetectBenchmark. Combining LSDB and GSDB raises mAcc to 87.2%, supporting the contribution of global spectral information. With the same dual-branch configuration, adding BCGF further increases mAcc to 92.2% and mAP from 93.2% to 96.9% (Table 6), supporting bidirectional calibration and adaptive fusion.

Our contribution therefore lies in a unified framework combining explicit neighborhood-context residuals, complementary local-subband and global-spectral modeling, and bidirectional adaptive integration. The controlled experiments support the contributions of the representation, complementary branches, and fusion module.

Additional concern: robustness and real-world generalization

On WildRF, our method achieves the highest mean accuracy among the compared methods, reaching 67.1% and exceeding the runner-up by 1.2 percentage points under the reported protocol.

To provide additional robustness evidence, we evaluated common post-processing operations on ForenSynths (Appendix F, Table 12):

Test conditionmAcc. (%)mAP (%)
No additional degradation84.893.5
JPEG compression (Q=85)82.391.6
Resizing (scale=0.75)80.090.1
Gaussian blur (sigma=0.5)84.294.0

The model retains at least 80.0% mAcc and over 90.0% mAP under all three conditions. JPEG compression reduces mAcc by only 2.5 percentage points, while Gaussian blur largely preserves accuracy. These experiments support the robustness of our method under the tested post-processing conditions and provide further evidence of its applicability to image dissemination scenarios.

Reviewer_TA1LOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a hierarchical dependency modeling framework for generalizable AI-generated image detection. It introduces Neighborhood Dependency Representation (NDR) to extract local residuals and reduce semantic interference, followed by Local Subband Dependency Branch (LSDB) and Global Spectral Dependency Branch (GSDB) to capture complementary spatial-spectral features. A Bidirectional Confidence Guided Fusion (BCGF) module adaptively integrates these representations. Trained on ProGAN-generated images, the method achieves competitive cross-generator performance on ForenSynths, AIGCDetectBenchmark, and WildRF, supported by component ablations and qualitative visualizations.

Critical strengths
  1. The paper addresses the important challenge of cross-generator generalization by focusing on dependency irregularities rather than generator-specific artifact appearances. The combination of neighborhood residuals, local–global spectral modeling, and adaptive fusion provides a coherent and practically motivated framework for learning transferable forensic features.

  2. Despite training on images from a single generator, the method achieves the highest macro-averaged accuracy across three benchmarks, covering diverse synthesis methods and real-world scenarios. The component ablations further demonstrate substantial contributions from NDR and BCGF, supporting the practical effectiveness of the proposed framework.

Critical weaknesses
  1. While the dependency-oriented perspective is reasonable, more direct evidence is needed to establish its advantages over conventional residual or spectral features and to validate NDR's claimed semantic suppression.
  2. LSDB and GSDB integrate established wavelet, Fourier, and attention operations, but their specific advantages in dependency modeling remain insufficiently isolated. Likewise, further analysis of BCGF's learned gates would help clarify the role of confidence-guided fusion in the observed improvements.
  3. The method outperforms DFFreq on only 5 of 17 AIGCDetectBenchmark subsets, with gains concentrated on SDXL and WFIR. Near-chance accuracy on CRN and IMLE further raises concerns about the consistency of the claimed transferable representations.
Questions for authors
  1. Could the authors demonstrate that dependency irregularities are more consistent across generators than conventional residual or spectral features, using controlled NDR comparisons that also assess semantic suppression?
  2. How do LSDB and GSDB capture dependencies beyond standard wavelet and Fourier processing? Could capacity-matched comparisons and fusion-weight analyses validate their contributions and BCGF's confidence-guided mechanism?
  3. How do the authors reconcile strong macro-averaged performance with uneven cross-generator gains, particularly those concentrated on SDXL and WFIR?
  4. Given the substantial discrepancies between ACC and AP on certain generators (e.g., CRN and IMLE), could the authors analyze whether these failures arise from insufficient forensic representations or score-calibration and threshold-transfer issues?
2 comments
AuthorsOpenReview ↗

Response to Reviewer TA1L

Title: Response to Reviewer TA1L

Comment:

We thank the reviewer for the constructive feedback. We address the four questions below.

Q1. NDR comparisons and semantic suppression

We added controlled comparisons that keep the downstream architecture and training settings fixed, replacing only the front-end representation. Appendix C.1, Table 8 reports mAcc./mAP (%):

RepresentationAIGCDetectBenchmarkForenSynthsWildRF
Mean-filter residual82.3 / 91.678.3 / 89.260.4 / 67.6
Median-filter residual86.1 / 95.185.0 / 94.664.2 / 67.5
NPR86.1 / 93.574.2 / 81.258.5 / 66.8
NDR92.2 / 96.984.8 / 93.567.1 / 71.0

NDR achieves the best results on AIGCDetectBenchmark and WildRF and near-best performance on ForenSynths. Unlike standard median filtering, NDR replaces the center pixel with zero before aggregation, making the reconstruction independent of its original value. The residual explicitly measures deviation from the surrounding context, highlighting potential synthesis-induced dependency irregularities.

For semantic suppression, we evaluate a fixed ImageNet-1K-pretrained Swin-T on the ImageNet-1K validation set (Appendix C.2, Table 9):

InputTop-1 (%)Top-5 (%)
RGB81.47495.768
NDR23.57640.800

The 57.898-point Top-1 reduction supports that NDR substantially suppresses class-discriminative information accessible to the pretrained classifier.

Q2. Branch modeling and adaptive fusion

We added a static fusion-weight control retaining both branches, bidirectional modulation, and the same number of trainable fusion parameters. Only the sample-dependent descriptors supplied to the final fusion gate are replaced with a constant vector. The weights remain learnable but are shared across samples. This differs from removing BCGF and directly combining branch outputs using a learnable scalar α.

Equal-weight macro-averages over the 17 AIGCDetectBenchmark subsets are:

SettingACC (%)AP (%)
Static learnable fusion86.2594.08
Full BCGF92.1896.87
Gain (percentage points)5.932.79

Full BCGF improves ACC on 16 of 17 subsets. This provides performance-based evidence that sample-dependent fusion improves detection beyond a globally shared mixing ratio.

We agree that capacity-matched controls would better isolate LSDB/GSDB’s contributions beyond standard wavelet/Fourier processing. These experiments remain incomplete due to time constraints; we will provide the results in a subsequent update.

Q3. Uneven gains across generators

We acknowledge that our method surpasses DFFreq on only 5 of 17 subsets and remains weaker on some generators. However, on SDXL and WFIR, accuracy improves from 64.2%/52.2% to 96.1%/88.9%, gains of 31.9/36.7 points. Our claim concerns mitigating severe generator-specific failures and improving overall generalization, rather than winning on every generator. Our method achieves the highest mean accuracy on all three benchmarks.

We also examined cropping without retraining. Averaging predictions from the four corners and center improves AIGCDetectBenchmark mAcc. from 92.18% to 94.36%, with gains of 6.46/6.56/7.32 points on SD v1.4/SD v1.5/Wukong. This suggests that useful evidence exists beyond the center and motivates improved spatial sampling (Appendix D, Table 10).

Q4. ACC–AP discrepancies on CRN and IMLE

Class-wise analysis shows 3.2% real-image accuracy and 100.0% fake-image accuracy on both subsets: errors primarily arise from reference images being classified as generated. Several compared methods also show low ACC on these subsets.

Training uses natural images from CNNDetection/ProGAN as real samples, whereas CRN/IMLE reference images originate from GTA-rendered scenes. This provenance difference introduces real-class distribution shift and may shift detection scores, causing fixed-threshold mismatch.

AP remains 74.7%/74.8%, indicating retained discriminative ability despite poor fixed-threshold accuracy. Consistently excluding these two subsets for all methods, our method achieves 90.9% ACC/97.0% AP on the remaining 11 subsets, still the highest among compared methods. We will add the class-wise results and discuss this domain shift and potential threshold mismatch in the revised manuscript.

AuthorsOpenReview ↗

Supplementary Response to Question 2

Thank you for this suggestion. We conducted additional parameter-matched comparisons and a static-fusion ablation to isolate the contributions of LSDB, GSDB, and BCGF. We kept NDR, the classification backbone, and the training protocol unchanged. The following results are macro-averaged across the 17 generator subsets of AIGCDetectBenchmark.

SettingTotal trainable parametersACC (%)AP (%)
Our default setting1,438,23092.1896.87
Replacing LSDB with a conventional wavelet-domain CNN1,438,23179.5586.89
Replacing GSDB with a conventional Fourier branch1,438,23088.9393.90
BCGF with static, learnable fusion weights1,438,23086.2594.08

For LSDB, the control preserves the wavelet transform, normalization, subband residual connection, and inverse wavelet reconstruction, while replacing the dependency-modeling operation with parameter-matched conventional convolutions. Our default setting improves ACC and AP by 12.63 and 9.98 percentage points, respectively, supporting the contribution of LSDB’s learned modeling structure beyond conventional wavelet-domain processing.

For GSDB, the control uses conventional FFT/IFFT and real–imaginary spectral convolutions with a matched branch parameter budget, while retaining the remaining modules. Our default setting improves ACC and AP by 3.25 and 2.97 percentage points, respectively, supporting the effectiveness of GSDB’s specific spectral modeling structure.

For BCGF, the static control retains bidirectional modulation and the same number of trainable fusion parameters. Only the sample-dependent descriptor supplied to the final fusion gate is replaced with a constant vector, yielding globally shared but learnable fusion weights. This differs from the existing ablation that removes BCGF entirely and combines the branch outputs using α and 1−α. Our default setting improves ACC and AP by 5.93 and 2.79 percentage points, respectively, supporting the benefit of adjusting the final fusion weights to individual inputs.

These controlled comparisons provide empirical support for the contributions of LSDB, GSDB, and sample-dependent fusion under closely matched parameter budgets.

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

This paper proposes a generalizable AI-generated image detection framework based on hierarchical dependency modeling. Specifically, the authors introduce a Neighborhood Dependency Representation (NDR) to suppress semantic interference and highlight irregularities in local pixel dependencies. Two complementary branches, the Local Subband Dependency Branch (LSDB) and the Global Spectral Dependency Branch (GSDB), are then developed to capture local and global dependency structures, respectively. Finally, a Bidirectional Confidence Guided Fusion (BCGF) module adaptively integrates these representations for real/fake classification.

Extensive experiments on ForenSynths, AIGCDetectBenchmark, and WildRF demonstrate competitive cross-generator generalization performance, achieving average accuracies of 84.8%, 92.2%, and 67.1%, respectively.

Critical strengths
  1. Important research problem and well-motivated approach. The paper addresses cross-generator generalization, which remains a significant challenge in AI-generated image detection. The motivation to capture transferable forensic evidence by modeling local and global dependency irregularities is reasonable and practically relevant.

  2. Reasonable methodological design with meaningful contributions. The proposed framework combines neighborhood dependency residual representations with local wavelet-based and global Fourier-based dependency modeling. Furthermore, the BCGF module employs bidirectional feature calibration and adaptive fusion to integrate complementary information. The overall architecture is well structured, and the individual components are logically connected.

  3. Competitive experimental performance. The proposed method achieves the highest mean classification accuracy among the compared methods on ForenSynths, AIGCDetectBenchmark, and WildRF. In particular, it achieves 92.2% mean accuracy on AIGCDetectBenchmark, outperforming recent representative approaches such as DFFreq. These results demonstrate the effectiveness of the proposed framework for cross-generator detection.

  4. Comprehensive ablation studies. The authors conduct component-wise ablation studies to investigate the contributions of different modules. According to Table 6, removing NDR and BCGF decreases the mean accuracy by 7.4 and 5.0 percentage points, respectively, indicating their importance to the overall performance. In addition, the experiments evaluate the individual contributions of LSDB and GSDB, as well as the influence of different neighborhood sizes. These results provide useful empirical support for the effectiveness of the proposed framework.

Critical weaknesses
  1. The performance variations across different generators deserve further analysis. Although the proposed method achieves strong overall mean accuracy, its performance varies across different generative models. For example, on AIGCDetectBenchmark, the proposed method significantly outperforms DFFreq on SDXL but performs worse on several other subsets, including ADM, Stable Diffusion v1.4/v1.5, and Wukong.

These results suggest that the effectiveness of the learned dependency representations may vary across different generation mechanisms. A more detailed analysis of these performance differences would help clarify the applicability and limitations of the proposed method across different generator families.

  1. The robustness evaluation under real-world image degradations could be more comprehensive. The proposed method achieves 67.1% mean accuracy on WildRF, outperforming the compared approaches. However, the improvement over the strongest baseline is relatively modest.

Although WildRF provides a meaningful evaluation under real-world social-media distribution shifts, the paper does not include controlled experiments investigating the effects of specific image degradation operations.

Since images distributed through online platforms often undergo compression, resizing, and other post-processing operations, additional experiments under JPEG compression, image resizing, and blurring would help provide a more comprehensive understanding of the robustness of the proposed neighborhood dependency representation.

  1. The analysis of the NDR representation could be further strengthened. The proposed NDR constructs residual features by subtracting a neighborhood-based median reconstruction from the original image. The existing ablation results demonstrate the effectiveness of NDR, and Table 7 further investigates the influence of neighborhood size.

However, the specific characteristics of the information preserved by NDR and its differences from conventional local filtering residuals could be analyzed in greater depth.

For instance, comparisons with standard high-pass filtering residuals or alternative local filtering strategies would help clarify the characteristics of NDR in preserving synthesis-related forensic information.

Additionally, the motivation for introducing an extra zero value in Equation (1), as well as its influence on neighborhood reconstruction, could be explained more clearly.

  1. Computational complexity and inference efficiency are not sufficiently evaluated. The proposed framework incorporates local-window attention, frequency-domain transformations, and bidirectional feature fusion. However, the paper does not provide a detailed analysis of model parameters, FLOPs, or inference time.

Although the training configuration and hardware environment are described, these details are insufficient to fully characterize the computational overhead of the proposed approach.

A comparison of computational complexity and inference efficiency with representative baselines would help assess the practical applicability of the proposed framework.

  1. The detection performance across different image regions requires further investigation. During evaluation, the proposed method consistently applies a 256×256 center crop to each input image. However, images generated by different models may vary considerably in resolution, content distribution, and local texture characteristics.

Relying exclusively on the central region may overlook synthesis-related artifacts present in other image regions, making it difficult to fully assess the spatial robustness of the proposed detector.

It would be useful to compare different evaluation strategies, such as center cropping, random cropping, and multi-region cropping, to investigate whether the proposed method can consistently capture forensic evidence across different spatial locations.

Questions for authors

Q1. Performance variations across generators. The proposed method significantly outperforms DFFreq on SDXL but shows lower accuracy on several other diffusion-based generators, including ADM and Stable Diffusion v1.4/v1.5. Could the authors provide further analysis of the factors contributing to these performance differences?

Q2. Characteristics of the NDR representation. Could the authors further clarify the differences between NDR and conventional local filtering residuals, potentially through additional comparative experiments? Moreover, what is the specific motivation for introducing an additional zero value in Equation (1)?

Q3. Robustness under image degradations. Could the authors provide additional evaluations under common image processing operations, such as JPEG compression, resizing, and blurring, to assess the stability of the proposed method under different image degradation conditions?

Q4. Computational efficiency. Could the authors report the parameter count, FLOPs, and inference time of the proposed framework and compare them with representative baselines to better characterize its computational efficiency?

Q5. Sensitivity to image cropping strategies. Since the evaluation protocol uses a fixed center crop, could the authors investigate the effects of different cropping strategies, including random cropping and multi-region cropping, to assess the consistency of detection performance across different image regions?

1 comment
AuthorsOpenReview ↗

Response to Reviewer CtBj

Thank you for the constructive review and for recognizing our motivation, method design, experimental performance, and ablation studies. We address your questions below. The added experiments are included in the revised manuscript's appendices.

Q1. Performance variations across generators

Performance variations may relate to preprocessing and training augmentation, and may also be partly affected by the cropping strategy raised in Q5. The additional experiments for Q5 show that, with unchanged model weights, multi-region cropping improves detection on ADM, SD v1.4/v1.5, and Wukong, indicating useful detection cues beyond the image center. We thank the reviewer for this suggestion, which motivates further optimization of preprocessing, augmentation, and spatial sampling to improve cross-generator detection.

Q2. NDR versus conventional residuals and the zero in Eq. (1)

We added controlled comparisons with the same downstream architecture and training settings, replacing only the front-end representation (Appendix C.1, Table 8). Entries are mAcc/mAP (%).

Front endAIGCDetectBenchmarkForenSynthsWildRF
Standard mean-filter residual82.3/91.678.3/89.260.4/67.6
Standard median-filter residual86.1/95.185.0/94.664.2/67.5
NPR86.1/93.574.2/81.258.5/66.8
NDR (ours)92.2/96.984.8/93.567.1/71.0

NDR achieves the best mAcc and mAP on AIGCDetectBenchmark and WildRF. Compared with the standard median-filter residual, the gains are 6.1/1.8 and 2.9/3.5 percentage points, respectively, while performance remains close to the best on ForenSynths.

The key to NDR is highlighting dependencies between the center pixel and its neighbors. Unlike standard median filtering, we set the center pixel to zero before median aggregation, making the reconstruction independent of its original value and more dependent on surrounding pixels. The difference between the original pixel and this reconstruction explicitly measures its deviation from neighborhood context, highlighting potential synthesis-induced dependency irregularities. These explanations and comparisons are included in Appendix C.1.

Q3. Robustness to image degradations

We added evaluations on ForenSynths under common post-processing operations (Appendix F, Table 12).

Test conditionmAcc (%)mAP (%)
No additional degradation84.893.5
JPEG (Q=85)82.391.6
Resizing (scale=0.75)80.090.1
Gaussian blur (σ=0.5\sigma=0.5)84.294.0

Across these tested conditions, mAcc remains at least 80.0% and mAP exceeds 90.0%. JPEG reduces mAcc by 2.5 percentage points, while mild blur largely preserves accuracy. These results support the framework's robustness under the tested post-processing conditions.

Q4. Computational efficiency

Appendix E, Table 11 reports efficiency and detection performance on AIGCDetectBenchmark. Our model has 1.438M parameters, requires an estimated 4.567 GFLOPs per 256 x 256 image, and achieves 437.56 FPS on an RTX 4090, with 92.2% mAcc and 96.9% mAP.

MethodInputParams (M)Est. GFLOPsFPSmAcc/mAP (%)
UFD224 x 224303.967162.438114.4977.4/91.1
NPR256 x 2561.4384.523996.7782.6/86.9
FreqNet256 x 2561.8508.224385.3378.2/85.3
SPAI256 x 256139.945659.86532.1260.7/68.2
CO-SPY224 x 224305.405165.902101.2585.6/95.0
MLEP256 x 2561.4414.639377.5985.3/91.7
DFFreq256 x 2561.4424.596502.6690.6/93.6
Ours256 x 2561.4384.567437.5692.2/96.9

Compared with NPR, our method improves mAcc/mAP by 9.6/10.0 percentage points with the same parameter count at the reported precision and only 0.044 additional GFLOPs per image, balancing detection performance and computational cost.

Q5. Sensitivity to cropping strategies

We evaluated 256 x 256 crops on AIGCDetectBenchmark with the same trained weights, without retraining (Appendix D, Table 10).

StrategymAcc (%)mAP (%)
Center crop92.1896.87
Random single crop92.61 ± 0.0597.22 ± 0.06
Four corners and center94.3698.05
Random five crops93.96 ± 0.0598.07 ± 0.01

Multi-crop inference averages regional prediction probabilities. Random strategies report mean ± sample standard deviation over five seeds; metrics are macro-averaged over generators. Random single cropping achieves comparable and slightly higher overall performance than center cropping. Four-corner-and-center aggregation improves mAcc by 2.18 percentage points, indicating useful evidence beyond the center. We thank the reviewer for this suggestion, which motivates further optimization of preprocessing, augmentation, and spatial sampling to improve cross-generator detection.

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

Related papers

Discussion (0)

Sign in to comment.