acceptodds
Withdrawn from ICLR 2027

Detector-Guided Fake Synthesis Enhancement for Generalizable AI-Generated Image Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

AI-generated image (AIGI) detectors often generalize poorly to unseen generators and image domains because they may exploit shortcut cues specific to the training distribution. Recent data-alignment methods reduce real-fake discrepancies but can introduce new pattern-specific artifacts through pixel-space synthesis. To address this limitation, we propose a feature-space fake synthesis framework that constructs additional fake-labeled training samples by interpolating aligned real and synthetic representations, avoiding additional pixel-level transformations while producing challenging fake features closer to the real distribution. We further introduce detector-guided adversarial optimization, where detector learning and fake-feature synthesis are alternated so that the evolving detector guides the generation of informative hard examples. To enable efficient adversarial optimization and pair-adaptive fake synthesis, we design a lightweight mixing coefficient generator that predicts a bounded interpolation coefficient for each real-fake pair. Additionally, we introduce a coefficient diversity constraint to mitigate generator collapse during adversarial training. Extensive experiments on nine benchmarks demonstrate improved generalization to unseen generators and image domains.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI-generated image detection, out-of-distribution generalization, shortcut learning, adversarial training

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_pb62OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a detector-guided feature synthesis framework to improve the cross-generator generalization of AI-generated image detection. To reduce shortcut learning caused by pixel-level synthesis artifacts, the method constructs hard fake features by interpolating aligned real and fake representations. A lightweight mixing coefficient generator is then optimized adversarially using detector feedback to produce more challenging, pair-adaptive fake features. Additionally, a diversity constraint is introduced to prevent coefficient collapse during training.

Critical strengths
  1. This paper is well organized and easy to follow.

  2. Sufficient discussions in the paper to help researchers understand the method: detailed comparison experiments and method discussions.

Critical weaknesses
  1. I have doubts about the core assumption of this paper that "the mixture of real and generated features still belongs to the generated category." Simple linear feature interpolation cannot guarantee that the mixed features correspond to a real generated image, let alone guarantee that it still retains reliable generation traces. Therefore, directly labeling all mixed features as Fake lacks sufficient theoretical basis and may introduce label ambiguity.
  2. The fairness of the comparative experiments is questionable. The authors directly used the pre-trained weights officially released by each comparative method, but these models were not obtained under the same dataset and training protocol, so the performance differences may not entirely come from the methods themselves.
  3. The authors emphasize that pixel space mixing may introduce additional artifacts, but this paper still relies on pixel space-aligned images constructed by DDA for training. This means that the method has not eliminated the bias that pixel-space data construction may bring, and there is still a certain gap between its core motivation and actual implementation.
  4. The current ablation experiments lack validation on different backbones. It is recommended that the authors further analyze the performance of this method under different feature extraction networks to prove that its effectiveness does not depend on specific pre-trained representations.
Questions for authors

Please refer to the Weaknesses section above for detailed questions and concerns.

  1. What justifies assigning the Fake label to interpolated real–fake features? How can the authors ensure that these features retain meaningful forensic traces rather than introducing label ambiguity?
  2. Since the compared methods use official checkpoints trained on different datasets and under different protocols, how can the authors ensure that the performance comparisons are fair?
  3. The paper argues that pixel-space synthesis introduces spurious artifacts, yet the training data still rely on DDA-generated aligned images. How does the proposed method avoid inheriting these artifacts and the associated biases?
  4. The experiments only use DINOv2 as the backbone. Can the authors clarify whether the proposed feature synthesis strategy is expected to generalize to other backbone architectures?
4 comments

Dear Reviewer #pb62, Thank you for taking the time to review our paper and for your constructive feedback. We hope the clarifications and additional results below address your concerns.


W1 & Q1. Justification for fake labels on mixed features

  • To explain which cues can survive interpolation, we consider the feature difference between aligned real and reconstructed images. For each pair, zm−zr=(1−λ)(zf−zr)z^m-z^r=(1-\lambda)(z^f-z^r). If a cue is expressed as a linear contrast between the endpoints, the mixed feature preserves the sign of that contrast relative to the real endpoint and scales its magnitude by 1−λ1-\lambda. The mixed feature can therefore expose the detector to a weaker version of the same cue.

  • We label the mixed features as fake to train the detector on weaker forensic cues. To limit label ambiguity, we set λmax⁡=0.8\lambda_{\max}=0.8, which keeps each mixed feature at least 20% of the original pairwise distance from its real endpoint. We also supervise the original real image and its reconstruction with their respective labels.

    In Table 3, the fixed-coefficient feature synthesis variant (λ0=0.65\lambda_0=0.65) achieves 86.10% mean ACC across nine benchmarks, compared with 82.64% for the original no-synthesis baseline.

W2 & Q2. Baseline comparison on matched training data

  • To control for training-data differences, we retrained NPR, UnivFD, SAFE, C2P-CLIP, AIDE, and DRCT on the same COCO-SD2 dataset as our method. It consists of real MS COCO images paired with their Stable Diffusion 2.0 VAE reconstructions. The AlignedForensics and DDA results in the original submission already come from models we trained on this dataset. FatFormer uses the official checkpoint because its training code was unavailable to us.

  • We evaluated all detectors on the same nine benchmarks. Mean is the unweighted average of the nine benchmark-level accuracies. AIGCDB denotes AIGCDetectionBenchmark.

    Table R1. Accuracy on nine benchmarks with matched training data

    MethodAIGCDBDRCT-2MWildRFForenSynthsEvalGENSynthbusterSynthWildXChameleonGenImageMean
    NPR50.1655.5252.2540.4731.0726.6750.0359.0450.8246.23
    UnivFD78.5386.8556.0067.4188.8762.3056.2561.2585.1271.40
    FatFormer74.7851.9555.6084.5248.9353.6253.3858.0361.0360.20
    SAFE65.3065.1656.9355.6061.1080.6354.6954.7369.3862.61
    C2P-CLIP66.1569.8559.3250.9553.8049.3654.2556.7374.6859.46
    AIDE50.5869.3158.6141.9537.6061.0660.3563.3855.1355.33
    DRCT78.7485.3458.8768.8373.9370.6560.7163.5181.5671.35
    AlignedForensics65.9595.3082.4151.7767.3178.2378.9770.2077.6074.19
    DDA86.9285.1878.1477.9173.7484.1575.1173.0383.0179.69
    Ours92.5296.2189.3786.8489.7690.9788.7774.2493.5689.14

    Our method achieves the highest reported accuracy on all nine benchmarks. Its mean ACC is 89.14%, which is 9.45 percentage points higher than DDA (79.69%), the strongest baseline trained on the same data. The accuracy advantage therefore persists when the training dataset is held fixed.

W3 & Q3. Pixel-mixing shortcuts and inherited biases

  • Pixel-blending artifacts can become shortcuts when blended images are labeled fake. To avoid introducing these cues during hard-negative synthesis, we encode each real image and its DDA-aligned reconstruction separately, then pass their interpolated features directly to the classifier.

  • To assess sensitivity to pixel blending, we compared fake detection rates on paired Recon and PixelMix inputs. Each detector was evaluated on 1,000 pairs of VAE reconstructions and their pixel-space blends with the source images. The gap is the fake detection rate on PixelMix minus that on Recon, in percentage points (pp).

    Table R2. Fake detection rates on Recon and PixelMix

    Detectorgap (pp)
    PixelMix-trained+27.3
    FeatureMix-trained+4.7

    The smaller gap for the FeatureMix-trained detector suggests that feature mixing reduces the detector’s reliance on pixel-blending cues as a shortcut for identifying fakes.

W4 & Q4. Transfer to a CLIP backbone

  • To test whether our framework transfers beyond DINOv2, we trained a CLIP variant on COCO-SD2 and evaluated it on the same nine benchmarks. UnivFD, C2P-CLIP, and DRCT use the matched-data results from Table R1. FatFormer uses its official checkpoint.

    Table R3. Accuracy of CLIP-based detectors across nine benchmarks

    MethodAIGCDBDRCT-2MWildRFForenSynthsEvalGENSynthbusterSynthWildXChameleonGenImageMean ACC (%)
    UnivFD78.5386.8556.0067.4188.8762.3056.2561.2585.1271.40
    FatFormer74.7851.9555.6084.5248.9353.6253.3858.0361.0360.20
    C2P-CLIP66.1569.8559.3250.9553.8049.3654.2556.7374.6859.46
    DRCT78.7485.3458.8768.8373.9370.6560.7163.5181.5671.35
    Ours (CLIP)86.6196.7680.7171.5586.9765.8380.0764.5190.8480.43

    Our CLIP variant achieves 80.43% mean ACC, 9.03 percentage points above the strongest baseline in this comparison. These results support the applicability of our feature synthesis framework to CLIP.

Thank you for your detailed reply, which resolved my question, so I am increasing my score.

Thank you for reading our response and updating your assessment. We are glad our reply addressed your question, and we appreciate your constructive feedback.

Reviewer_o322OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a training-time feature-space augmentation strategy for AI-generated image detection. It interpolates deep representations of semantically aligned real and fake image pairs and labels the interpolated features as fake. A lightweight MLP predicts a bounded, pair-specific mixing coefficient. The detector and coefficient generator are updated alternately: the detector learns from the synthesized features, while the generator seeks mixtures that the current detector finds difficult to classify. A batch-level coefficient-diversity regularizer discourages all mixing coefficients from collapsing to the same value. The additional generator is discarded at inference. Experiments cover nine benchmarks, with ablations of feature synthesis, adversarial optimization, and diversity regularization.

Critical strengths
  1. Practical training-only augmentation. Synthesizing hard examples in feature space is straightforward to integrate into a detector and does not require an additional generation module at inference. It also avoids performing another pixel-space synthesis operation during augmentation.

  2. Broad evaluation and useful component ablations. The experiments span nine benchmarks and report ACC, AUC, and AP. Table 3 shows that feature synthesis improves the detector under the authors' setup, with further gains from adversarial optimization and coefficient-diversity regularization. These controlled comparisons support the effectiveness of the overall training procedure.

Critical weaknesses
  1. The incremental technical contribution and the necessity of pair-adaptive mixing are not sufficiently established. Feature interpolation, adversarial hard-example generation, and diversity regularization are familiar techniques. The potentially distinctive element here is the learned mixing coefficient conditioned on each aligned real/fake pair, but the paper does not sufficiently separate its contribution from simpler alternatives. Equation (3) and the component ablation already include a fixed-coefficient feature-synthesis baseline; nevertheless, they do not provide decisive comparisons against a fixed coefficient at the upper bound λ_max, coefficients sampled randomly from the same interval, or coefficients predicted for one pair and reassigned to another. This distinction is important because the adversarial objective in Eq. (9) tends to favor larger real-feature contributions when the detector correctly orders a pair's real and fake endpoints, a tendency also acknowledged in Section 3.4 and Figure 2. Moreover, the detector loss in Eq. (5) uses one real term and two fake-labeled terms per pair, so the effects of adding hard examples should be disentangled from a change in class/loss weighting. Without these controls, the evidence more convincingly supports an effective combination of training techniques than the necessity of the proposed pair-adaptive optimization mechanism.

  2. Systematic arithmetic inconsistencies in Table 17. The caption states that the Mean AP column is the unweighted average of the nine benchmark-level values. Recomputing these values shows that the reported mean is inconsistent for all nine baseline methods, while the mean for the proposed method is consistent. For example, NPR's nine entries average to approximately 52.31%, not the reported 50.31%; DRCT averages to approximately 80.13%, not 83.89%; and DDA averages to approximately 92.58%, not 93.38%. The discrepancies are much larger than rounding errors. The authors need to explain how this column was produced, correct the table, and verify all associated comparative statements.

Questions for authors
  1. What is the specific technical advance beyond existing feature interpolation and adversarial augmentation? Under the same backbone, data, optimization budget, and class weighting, how does the full method compare against fixed λ = λ_max and random mixing within [λ_min, λ_max]? Does performance decrease if learned coefficients are randomly reassigned across pairs?
  2. Please specify the classifier architecture, the coefficient-generator MLP architecture, the number of aligned training pairs, and the coefficient used in the existing fixed-mixing ablation. Were λ_min, λ_max, and the diversity weight γ chosen on a held-out validation split, rather than on any of the reported evaluation sets? Can the authors also provide a class-weight-matched ablation for Eq. (5)?
  3. How were the Mean AP values in Table 17 calculated? Please provide a corrected table and confirm that the methods were evaluated on the same subsets and under consistent aggregation procedures.

I would reconsider my recommendation if the authors can adequately address these concerns.

4 comments

Dear Reviewer #o322, Thank you for taking the time to review our paper and for your constructive feedback. We hope the clarifications and additional results below address your concerns.


W1 & Q1. Technical contribution of pair-adaptive mixing

  • To account for differences in pair difficulty, we adapt the mixing strength using detector feedback. A single mixing coefficient can produce an easy negative for one aligned real–fake pair and an ambiguous mixture for another. We use feedback from the current detector to adjust the interpolation strength for each pair. This lets us generate hard negatives that challenge the detector as it learns, giving it practice distinguishing subtler real–fake differences. The upper bound preserves a minimum contribution from the fake feature, and diversity regularization discourages the coefficients from collapsing to the upper bound (Fig. 2). This gives the generator room to adapt the mixing strength to each pair in response to detector feedback, exposing the detector to hard negatives with varying levels of fake-feature contribution.

  • To test the contribution of pair-adaptive mixing, we compare it with controls that vary how coefficients are assigned. We compare the full method with fixed upper-bound mixing (λ=0.8\lambda=0.8) and random mixing (independent uniform sampling from [0.5,0.8][0.5,0.8] for each pair). Both controls use the same DINOv2-LoRA backbone, training data, detector optimization settings, training budget, and classification loss weights as the full method. We also shuffle the recorded coefficients across pairs within each mini-batch, preserving their empirical distribution. The feature-synthesis baseline in Table 3 uses the midpoint λ0=0.65\lambda_0=0.65.

  • To isolate the effect of feature synthesis, we match the real-to-fake loss weights in the no-mixing control. For the no-mixing control, we count the classification loss once for each real image and twice for its reconstruction, then average the three terms. This matches the full method's total classification loss weight for each class.

    Table R1. Accuracy by mixing strategy

    SettingAIGCDBDRCT-2MWildRFForenSynthsEvalGENSynthbusterSynthWildXChameleonGenImageMean ACC (%)
    Feature synthesis (Table 3)90.8695.9986.2678.7289.4688.3286.4768.0190.8086.10
    All coefficients fixed at λmax⁡\lambda_{\max}90.0294.2585.2585.0587.5088.3083.2369.7090.3585.96
    Random coefficients89.6395.7387.6380.8585.2389.5885.7970.0090.8186.14
    Shuffled learned coefficients91.3096.0686.8884.0581.1089.7183.5870.2193.8686.31
    No feature mixing92.3497.4782.3586.0557.7085.8878.0268.2991.9382.23
    Full method (Table 3)92.5296.2189.3786.8489.7690.9788.7774.2493.5689.14

    The shuffled control uses the same recorded coefficient values in each mini-batch and changes only which pair receives each coefficient. Its lower accuracy relative to the full method in Table 3 is consistent with a benefit from pair-specific coefficient assignment.

Q2. Architecture, training, and model selection

  • The classifier maps the 1,024-dimensional feature to one logit. A sigmoid converts the logit to a fake probability. The coefficient generator concatenates the two endpoint features and passes them through a 2048→256→64→12048\to256\to64\to1 MLP. Both hidden layers use LeakyReLU with a negative slope of 0.2. The scalar output passes through a sigmoid and is rescaled to [λmin⁡,λmax⁡][\lambda_{\min},\lambda_{\max}].

  • We trained on 118,287 aligned real–fake pairs. Each pair consists of a real MS COCO image and its Stable Diffusion 2.0 VAE reconstruction. The fixed-mixing ablation in Table 3 uses λ0=0.65\lambda_0=0.65 for every pair.

  • We selected the coefficient bounds on GenImage. We varied one bound at a time (Fig. 3) and selected λmin⁡=0.5\lambda_{\min}=0.5 and λmax⁡=0.8\lambda_{\max}=0.8. We set γ=5\gamma=5 by comparing the magnitudes of the batch-wise coefficient standard deviation and the adversarial loss. The other eight benchmarks were used only for evaluation.

  • We used the final saved checkpoint from one training run per configuration. All final checkpoints were saved at the same training step.

  • To control for class weighting in Eq. (5), we trained a detector without feature mixing using the same 1:2 real-to-fake loss weights. For each pair, we counted the classification loss once for the real image and twice for its reconstruction, then averaged the three terms. The full method achieves higher mean ACC than the no-mixing control, indicating that feature mixing improves detection under matched class weights.

    Table R2. Feature mixing with matched class weights

    SettingAIGCDBDRCT-2MWildRFForenSynthsEvalGENSynthbusterSynthWildXChameleonGenImageMean ACC (%)
    No feature mixing92.3497.4782.3586.0557.7085.8878.0268.2991.9382.23
    Full method (Table 3)92.5296.2189.3786.8489.7690.9788.7774.2493.5689.14

W2 & Q3. Mean AP correction and evaluation protocol

Thank you for carefully checking Table 17. We found an error in our calculation of Mean AP. Some benchmark-level AP values may have been stored as text in the spreadsheet and skipped by the averaging formula. We recalculated each mean using all nine benchmark scores and will update Table 17 and the related comparisons in the revision.

Corrected Table 17. Average precision (%) across nine benchmarks

MethodAIGCDBDRCT-2MWildRFForenSynthsEvalGENSynthbusterSynthWildXChameleonGenImageMean AP
NPR48.4153.7969.1547.5772.9945.1946.0940.9146.7252.31
UnivFD87.1386.2054.1191.7588.8260.2951.3747.6882.5472.21
FatFormer79.6346.5077.8896.2278.7755.8158.2557.8068.0368.76
SAFE49.8757.6570.6645.4374.0657.3548.5150.6456.1356.70
C2P-CLIP85.8085.3469.1797.4787.7171.3155.8440.1982.2275.01
AIDE59.3674.7059.1044.7488.8052.2053.1951.7572.4661.81
DRCT90.4296.1259.3688.5797.7274.5354.0866.4093.9380.13
AlignedForensics77.5199.7191.9365.2497.9578.1987.7780.2188.5585.23
DDA95.3496.8488.8394.6498.1493.0988.5281.1396.6592.58
Ours97.9399.2895.5795.8098.6397.4894.3781.3598.3395.42

All methods were evaluated on the same benchmark subsets using the same AP aggregation procedure. Our method ranks first on seven of the nine benchmarks and achieves the highest mean AP of 95.42%, 2.84 percentage points ahead of DDA (92.58%). We will correct Table 17 and update the associated comparisons in the revised manuscript.

Dear Reviewer #o322,

We appreciate the time and care you put into reviewing our paper. We hope our response helped clarify the points you raised. If there is anything you would like us to elaborate on, we would be glad to do so.

Thank you for addressing my comments; after considering your response, I have decided to keep my original rating.

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

The paper augments an aligned real/fake training set by interpolating between paired DINOv2 features and labeling the interpolated features as fake. A small coefficient generator is trained adversarially against the detector so that these features are difficult to classify, while a batch-level diversity term discourages the coefficients from collapsing. The detector is adapted with LoRA, and the coefficient generator is discarded at inference. The reported evaluation spans nine benchmarks and shows improvements in mean accuracy and AUC over the compared methods. This is a practical idea with encouraging results; however, it remains unclear whether the current evidence genuinely isolates the learned mixing mechanism and whether the baseline comparison is documented rigorously enough to substantiate the reported gains.

Critical strengths
  1. The approach is simple enough to be useful. It adds hard training examples without another image-generation stage, and it incurs no coefficient-generator cost at inference. Appendix B makes the alternating optimization reasonably clear, including detaching coefficients during detector updates and recomputing endpoint features for generator updates. Those details are helpful for understanding what is actually optimized.
  2. The empirical coverage is a real strength. The paper evaluates one trained model on both curated and in-the-wild benchmarks and supplies threshold-independent results in the appendix. Table 10 reports a mean AUC of 95.54% versus 92.78% for DDA, so the reported benefit is not confined to accuracy at a single threshold. More importantly for attribution, Table 3 improves from 82.64% without feature synthesis to 89.14% with the full method. That within-setup comparison is more informative than the headline comparison to separately released checkpoints.
Critical weaknesses
  1. The ablations do not yet show that pair-adaptive adversarial mixing is the essential ingredient. Table 3 is useful, but the fixed mixing coefficient for its feature-synthesis-only row is not specified, and there is no comparison to random coefficients over the same interval or to shuffled learned coefficients. The paper explicitly observes that the adversarial objective pushes coefficients toward the upper bound. Maximizing their standard deviation prevents a constant solution, but does not itself establish that the resulting assignments reflect meaningful pair difficulty. The largest gain after fixed mixing comes from diversity regularization (2.28 points), while adversarial optimization alone adds 0.76 points. I would therefore want a matched-distribution random or shuffled control before attributing the improvement to detector-guided pair adaptation.
  2. Eq. (5) gives two fake-labeled terms for every real term. A no-mixing control with matched class weighting would help distinguish better features from a changed training balance. Since the endpoints are DDA-aligned images, the method avoids additional pixel operations but does not by itself show that existing reconstruction or alignment shortcuts have disappeared.
Questions for authors
  1. What fixed coefficient is used in the feature-synthesis-only ablation? How does the full method compare with random coefficients over the same bounds, shuffled coefficients preserving the learned distribution, and a no-mixing baseline with matched real/fake loss weights?
  2. How were the coefficient bounds, diversity weight, and final checkpoint selected? Were GenImage or any of the other eight evaluation benchmarks used for selection?
3 comments

Dear Reviewer #Uoux, Thank you for taking the time to review our paper and for your constructive feedback. We hope the clarifications and additional results below address your concerns.


W1. Contribution of pair-adaptive mixing

  • To test the contribution of pair-adaptive mixing, we compared the full method with fixed, random, and shuffled coefficients. We added two controls for pair-adaptive mixing. The random control independently samples a coefficient for each pair from the full method's range [0.5,0.8][0.5,0.8] using a uniform distribution. The shuffled control reassigns learned coefficients across pairs within each mini-batch, preserving the exact coefficient distribution. We also tested fixing every coefficient at λmax⁡\lambda_{\max} to examine the limiting case of coefficient collapse. The feature-synthesis-only setting in Table 3 uses the midpoint λ0=0.65\lambda_0=0.65 for all pairs.

    Table R1. Accuracy by mixing coefficient strategy

    SettingAIGCDBDRCT-2MWildRFForenSynthsEvalGENSynthbusterSynthWildXChameleonGenImageMean ACC (%)
    Feature synthesis (Table 3)90.8695.9986.2678.7289.4688.3286.4768.0190.8086.10
    Random coefficients89.6395.7387.6380.8585.2389.5885.7970.0090.8186.14
    Shuffled learned coefficients91.3096.0686.8884.0581.1089.7183.5870.2193.8686.31
    All coefficients fixed at λmax⁡\lambda_{\max}90.0294.2585.2585.0587.5088.3083.2369.7090.3585.96
    Full method (Table 3)92.5296.2189.3786.8489.7690.9788.7774.2493.5689.14

    The full method achieves 89.14% mean ACC, outperforming the random, shuffled, and fixed-λmax⁡\lambda_{\max} controls by 3.00, 2.83, and 3.18 percentage points, respectively. Shuffling preserves the learned coefficient distribution and changes the pair assignments, so the drop in accuracy supports the benefit of pair-adaptive mixing.

W2. Class weighting and inherited shortcuts

  • To control for class weighting, we trained a detector without feature mixing using the same 1:2 real-to-fake loss weights. For each pair, we averaged three loss terms, counting the real-image loss once and the reconstruction loss twice.

    Table R2. Feature mixing with matched class weights

    Training settingAIGCDBDRCT-2MWildRFForenSynthsEvalGENSynthbusterSynthWildXChameleonGenImageMean ACC (%)
    No feature mixing92.3497.4782.3586.0557.7085.8878.0268.2991.9382.23
    Full method92.5296.2189.3786.8489.7690.9788.7774.2493.5689.14

    With class weights matched, the full method achieves 6.91 percentage points higher mean ACC than the no-mixing control, supporting the contribution of feature synthesis.

  • To assess sensitivity to pixel-mixing shortcuts, we measured the change in fake detection rate from Recon to PixelMix. We evaluated each detector on 1,000 matched pairs of VAE reconstructions (Recon) and their pixel-space blends with the source images (PixelMix) under its respective test setting. We define the gap as the fake detection rate on PixelMix minus that on Recon, in percentage points (pp).

    Table R3. Change in fake detection rate from Recon to PixelMix

    Detectorgap (pp)
    PixelMix-trained+27.3
    FeatureMix-trained+4.7

    The smaller gap for the FeatureMix-trained detector suggests that feature mixing reduces the detector’s reliance on pixel-blending cues as a shortcut for identifying fakes.

Q1. Fixed coefficient and control comparisons

  • For the feature-synthesis ablation in Table 3, we used λ0=0.65\lambda_0=0.65 for every pair.

  • To assess the contributions of feature mixing and pair-specific coefficient assignment, we compare the full method with three controls. The random control samples uniformly from [0.5,0.8][0.5,0.8], the shuffled control preserves the learned coefficient distribution while reassigning coefficients across pairs, and the no-mixing control matches the full method's 1:2 real-to-fake loss weights. Table R4 reports ACC for each of the nine benchmarks and their mean. Δ\Delta is each control's mean ACC minus that of the full method, in percentage points.

    Table R4. Accuracy of the full method and controls

    SettingAIGCDBDRCT-2MWildRFForenSynthsEvalGENSynthbusterSynthWildXChameleonGenImageMean ACC (%)Δ vs. full
    Full method (Table 3)92.5296.2189.3786.8489.7690.9788.7774.2493.5689.14—
    Random coefficients89.6395.7387.6380.8585.2389.5885.7970.0090.8186.14-3.00
    Shuffled learned coefficients91.3096.0686.8884.0581.1089.7183.5870.2193.8686.31-2.83
    No feature mixing92.3497.4782.3586.0557.7085.8878.0268.2991.9382.23-6.91

Q2. Hyperparameter and checkpoint selection

  • We selected the coefficient bounds based on GenImage accuracy. We varied one bound at a time (Figure 3) and selected λmin⁡=0.5\lambda_{\min}=0.5 and λmax⁡=0.8\lambda_{\max}=0.8. The selected configuration achieved 93.56% ACC on GenImage. The lower bound ensures a minimum real-feature contribution to make the mixed samples harder, and the upper bound preserves a minimum fake-feature contribution to limit label ambiguity.

  • We set γ=5\gamma=5 to balance the diversity term and the adversarial loss. We based this choice on the relative magnitudes of the batch-wise standard deviation of the mixing coefficients and the adversarial loss during training.

  • We used the final saved checkpoint from a single training run per configuration. All checkpoints were saved at the same training step. We evaluated the selected checkpoint on all nine benchmarks. The eight benchmarks other than GenImage were used only for evaluation.

Dear Reviewer #Uoux,

Thank you again for your careful review and constructive feedback. We hope our response and the additional results were helpful. If you have any further questions, we would be happy to address them.

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

Related papers

Discussion (0)

Sign in to comment.