acceptodds
Under review as a conference paper at ICLR 2027

Asymmetry-Guided Iterative Adaptation for Generalizable AI-Generated Image Detection

Abstract

The continual emergence of new generative models challenges AI-generated image detectors whose forensic cues may fail beyond known sources. Empirical analysis indicates that target images initially predicted as synthetic provide useful signals for recognizing missed images from the same source, allowing partial cross-generator generalization to seed same-source adaptation. Based on this insight, we propose an iterative framework that repeatedly adapts with synthetic predictions to expand detection coverage. The framework uses asymmetric supervision: predicted-synthetic target images provide source-specific signals, while trusted real-image replay supplies real supervision without reinforcing missed synthetic images as real and suppresses false positives during iterative expansion. To overcome the limited coverage of initial synthetic predictions on difficult source distributions, we further introduce stochastic exploration, provisionally labeling a small subset of the remaining target images as synthetic while replay constrains noisy updates. Exploration accelerates recall expansion and improves final recall on difficult sources. Extensive experiments across unseen generators and diverse synthesis paradigms demonstrate state-of-the-art performance, outperforming recent methods by a substantial margin.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI-generated image detection, generalizable detection, iterative adaptation, asymmetric supervision

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_NRayOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes an asymmetry-guided iterative adaptation framework for AI-generated image detection, addressing limited generalization to unseen generators and the resulting false negatives. Its central observation is that initially detected synthetic images contain source-specific cues that can help recover other missed images from the same source. In contrast, using target images predicted as real as real-labeled training samples can reinforce errors associated with undetected synthetic images. At each iteration, the framework assigns synthetic pseudo-labels to target images predicted as synthetic and to a randomly sampled subset of the remaining images, enabling exploration beyond the detector’s current coverage. These samples are used jointly with independent, trusted real-image replay to optimize a binary cross-entropy objective, constraining false positives and decision-boundary drift caused by noisy pseudo-labels. The updated detector then re-evaluates the target collection, and adaptation proceeds until the growth in synthetic predictions saturates, subject to minimum and maximum iteration limits. The method is evaluated on 13 benchmarks. Across the 11 primary benchmarks, it improves average balanced accuracy from 87.2% to 98.5% on held-out evaluation subsets. Its main contribution is to turn partial cross-generator generalization into source-specific iterative adaptation while avoiding the reinforcement of false negatives through asymmetric supervision. These results are obtained through adaptation on unlabeled target data with an external trusted real-image pool, rather than zero-shot inference with a fixed detector.

Critical strengths

1.Well-motivated empirical insight. The preliminary experiments identify an important phenomenon: initially detected synthetic images provide useful cues for recovering missed images from the same source, whereas training on target images pseudo-labeled as real can reinforce existing false negatives. These findings provide a clear empirical basis for the proposed framework. 2.Thoughtful asymmetric supervision design. The framework obtains synthetic supervision from target data and real supervision from an independent, trusted real-image pool. This avoids repeatedly treating missed synthetic images as real during adaptation. Trusted real-image replay further helps control false positives and decision-boundary drift. 3.Strong empirical results and broad evaluation. The method achieves strong results across the reported benchmark comparisons. The evaluation spans 13 benchmarks covering multiple generation paradigms, in-the-wild images, and common post-processing operations. Component ablations, class-wise results, and analyses of difficult sources provide further evidence for the effectiveness of the proposed design.

Critical weaknesses

1.Insufficient characterization of noise introduced by stochastic exploration. The method assigns synthetic labels to randomly sampled images from the remaining pool, which may include genuine real images. Although trusted real-image replay is intended to constrain this label noise, it remains unclear under what conditions it can prevent the detector from learning spurious associations between target-domain characteristics and the synthetic class. This concern is particularly relevant when real images dominate the target collection. Further analysis under varying real-to-synthetic ratios would help establish the reliability of this mechanism. 2. Limited reliability for scarce or underrepresented synthetic sources. The effectiveness of iterative recall depends on both the number and the relative proportion of images from a difficult source. Figure 7 shows that gains can be unstable, and adaptation can underperform the initial detector in some low-count settings, such as those with only 10 images from the difficult source. The authors attribute this behavior to insufficient same-source evidence for learning a stable source-specific pattern. This limits the method’s applicability to emerging generators represented by only a few images in the target collection. 3.Dependence on the coverage of the trusted real-image pool. Trusted real-image replay draws on multiple datasets and additional web photographs and is an important component of stable adaptation. However, its ability to control false positives remains unclear when target real images differ substantially from the replay distribution, such as images from specialized acquisition devices or domain-specific imagery. A more detailed analysis of sensitivity to replay-pool size, composition, and distributional coverage would help clarify this dependency.

Questions for authors

1.Robustness to exploration-induced label noise. Stochastic exploration may assign synthetic labels to genuine real images in the target collection. Could the authors clarify under what conditions trusted real-image replay can effectively constrain the resulting label noise? Reporting synthetic-image recall and real-image false-positive rates across different target real-to-synthetic ratios and exploration budgets would help establish the robustness of this mechanism. 2.Construction and sensitivity of the trusted real-image replay pool. Trusted real-image replay appears central to stable adaptation. Could the authors provide the total pool size, the proportions contributed by each source, and details of the authenticity screening and deduplication procedures? Appendix A specifies that each round uses an equal number of replay real images and pseudo-labeled synthetic images; reporting the actual sample counts for representative benchmarks would further clarify the data requirements. Ablations on replay-pool size and source composition, together with an evaluation under a distribution mismatch between replay and target real images, would help characterize the method’s dependence on real-image coverage.

2 comments
AuthorsOpenReview ↗

Response to Reviewer NRay (1/2)

We sincerely thank the reviewer for the constructive comments. We have carefully addressed the concerns regarding exploration-induced label noise, scarce and underrepresented synthetic sources, and the construction and coverage of trusted-real replay by providing further clarifications and conducting additional experiments.

Response to W1 / Q1: Robustness to exploration-induced label noise

  • W1.1: Counteracting false positives with trusted real supervision. To verify the counteracting effect of trusted real replay, we design a controlled binary classification experiment with SDv1.4 images as fake and ImageNet images as real. We replace a fraction of the fake-class training images with ImageNet images while retaining the fake label, and examine how these injected false positives affect classification. Even when half of the fake-labeled set consists of real images, overall accuracy remains 99.80%, fake recall is 99.73%, and FPR is 0.13%, showing that trusted real supervision effectively counteracts such noise. In our framework, the number of randomly explored samples is bounded by the model-predicted fake count, so exploration contributes at most half of the pseudo-fake pool and the FP ratio during adaptation stays controllable. The remaining false positives are therefore within the range that trusted real replay can effectively constrain.
FP ratio %  Overall accuracy %  Fake recall %  FPR % 
0 99.9599.900.00
5 99.8099.630.03
10 99.3298.730.10
20 99.9299.830.00
50 99.8099.730.13
  • W1.2: Constraining false positives across rounds. The round-wise results on real benchmarks show a consistent trend that FPR remains low or decreases as adaptation proceeds on GenImage, AIGIBench, and Chameleon. Together with the controlled experiment in W1.1, these results indicate that trusted real replay continues to constrain false positives throughout iterative adaptation, rather than allowing them to accumulate over rounds.
Benchmark  Overall accuracy % by round  FPR % by round 
GenImage  88.25 → 94.81 → 96.77 → 98.64 → 99.24  0.29 → 0.08 → 0.08 → 0.25 → 0.19 
AIGIBench  75.38 → 88.14 → 97.14 → 97.67 → 97.98  3.50 → 1.17 → 1.12 → 0.83 → 0.50 
Chameleon  85.50 → 94.00 → 96.17 → 98.54 → 99.16  2.67 → 1.83 → 1.17 → 0.67 → 0.25 
  • W1.3: Robustness to the fake ratio in the target collection. We study whether the proportion of fakes in the target collection affects adaptation, especially when real images dominate. We fix GenImage fake images and vary the target fake:real ratio from 1:10 to 1:100, using random exploration whose size matches the predicted-fake count. Our sampling strategy limits the influence of this ratio in two ways. When few fakes are predicted, the collection likely contains fewer fakes, so the corresponding exploration budget also shrinks, reducing the noise introduced by labeling real images as fake. When many fakes are predicted, the exploration budget increases accordingly, allowing the framework to mine more difficult samples. As shown below, fake recall stays high and FPR remains low across all ratios, with accuracy above 98% even at 1:100.
Target fake:real  Fake recall %  FPR %  Balanced accuracy % 
1:10 98.120.1398.99
1:20 98.380.0899.15
1:50 98.250.1599.05
1:100 97.980.1198.94

Response to W2: Scarce and underrepresented synthetic sources

Our adaptation learns a source-specific within-class pattern from the predicted fakes to recall more fakes of the same source. With too few samples, the predicted fakes cannot form a stable pattern, leading to unstable adaptation. In practice, our framework is deployed in a dual-track manner: the online detector keeps serving while images accumulate, and the collection threshold that triggers offline adaptation can be set according to the desired stability, so a higher threshold delays adaptation until enough evidence is gathered. This instability is transient rather than persistent: as shown in Figure 7 of the paper, once a difficult source grows from 10 to 50 images, adaptation already yields largely stable positive gains across various fake proportions, and these gains increase further with more samples. In real deployment, new generators typically spread quickly once released; for instance, GPT-Image-2 first appear only in small numbers online but expand rapidly within a short period, so the accumulated data quickly reach the level required for stable adaptation.

AuthorsOpenReview ↗

Response to Reviewer NRay (2/2)

Response to W3 / Q2: Trusted-real pool construction, distribution mismatch, and pool size

  • W3.1: Stable anchor effect of trusted real supervision under distribution mismatch. We fix GenImage fake images and vary the target-real and replay-real sources to examine trusted real supervision under distribution mismatch. The cross-source results below show that replay can provide effective real-class constraints under distribution mismatch between target and replay images, supporting the stable anchor effect of trusted real supervision across differing source distributions. This effect is further enabled by the asymmetry of data collection: unknown fake images cannot be obtained in advance, whereas real images can be broadly pre-collected and augmented across scenes, contents, acquisition conditions, and processing pipelines, which improves the coverage of the real reference set and reduces the impact of target-real shift.
Target real Replay real  Fake accuracy  Real accuracy  Balanced accuracy 
ImageNet Open Images 97.6599.5098.58
MS COCO ImageNet 98.47100.0099.24
CC12M LAION 98.1397.3897.75
Open Images CC12M 97.1398.8898.00
  • W3.2: Sensitivity to replay-pool size. We vary the replay-pool size in multi-round adaptation on GenImage and AIGIBench. For efficient experiments, we randomly sample about 8,000 fake test images from GenImage and about 25,000 fake test images from AIGIBench. When the number of replay real images required by an update exceeds the available pool, replay draws are resampled with replacement. Small pools provide insufficient real constraints, while a sufficiently large pool supplies supervision with enough diversity for a stable counteracting effect, and the required pool capacity therefore depends on the scale of the target collection, with diminishing gains once sufficient coverage is available. Table entries are GenImage / AIGIBench.
Replay-pool size  Fake accuracy %  Real accuracy %  Balanced accuracy % 
1,000  99.50 / 99.90  80.94 / 28.58  90.22 / 64.24 
2,000  99.31 / 99.94  96.94 / 61.60  98.13 / 80.77 
5,000  99.25 / 99.96  99.06 / 91.48  99.16 / 95.72 
10,000  99.06 / 99.64  99.50 / 98.38  99.28 / 99.01 
20,000  98.56 / 99.60  99.69 / 99.10  99.13 / 99.35 
  • W3.3: Replay requirements on representative benchmarks. GenImage contains about 48k fake test images and AIGIBench contains about 112.5k fake test images. In each update, random exploration samples as many images as the model-predicted fakes, and trusted replay real images are used in the same number as the resulting pseudo-fakes. We report the replay requirements for the first three rounds, during which the predicted-fake counts grow fastest, and the corresponding replay requirements grow accordingly. Table entries are GenImage / AIGIBench.
Update round  Predicted-fake images  Exploration images  Required replay-real images 
1  35,346 / 65,849  35,346 / 65,849  70,692 / 131,698 
2  38,972 / 88,875  38,972 / 88,875  77,944 / 177,750 
3  42,990 / 90,576  42,990 / 90,576  85,980 / 181,152 
  • W3.4 Replay-pool construction: We build the trusted real replay pool by randomly drawing 50k images from each source dataset, so that each source contributes equally and the pool covers diverse scenes, contents, acquisition conditions, and processing pipelines. Regarding authenticity screening, candidate images are restricted to established, human-verified real-image datasets and publicly hosted photographs, and any samples whose provenance could not be confirmed are excluded. Deduplication is performed both within the pool and against all evaluation benchmarks, using ID matching of the datasets and image metadata, so that no replay image overlaps with the target collections used for adaptation or testing.
Reviewer_QvyLOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper considers the generalization capability of AI-generated image detectors when encountering previously unseen generation sources. The paper observes that if a synthetic image from an unseen source is correctly detected, this information can be leveraged to identify other synthetic images originating from the same source. Building on this insight, the paper proposes an iterative test-time adaptation framework that utilizes predicted synthetic images as "pseudo-labeled" training samples. In this approach, pseudo-labels are assigned exclusively to synthetic images, while the real-image class is supervised using reliable, authentic images. Furthermore, the work introduces a stochastic exploration mechanism: by randomly selecting additional target images and temporarily treating them as synthetic, the detector’s coverage of complex generation sources is enhanced. The detector undergoes multiple rounds of updates until the set of predicted synthetic images ceases to grow significantly. Experimental results across 13 benchmark datasets demonstrate that this method brings improvements over baselins.

Critical strengths
  • Useful empirical observation. The preliminary experiments in Figure 2 show that synthetic predictions from the target source are more useful for adaptation than samples from other generators. This is a reasonable motivation for the iterative adaptation framework.

  • The asymmetric supervision strategy is well motivated. Using predicted-real images as real pseudo-labels can reinforce the detector’s existing false negatives. The authors address this by using independently collected real images for supervision. The comparison in Table 5 supports this design choice.

  • The experimental results are good. The method improves average balanced accuracy from 87.2% to 98.5% across eleven benchmarks, with improvements on both synthetic and real images. Furthermore, the experiment includes useful analysis of the adaptation process.

Critical weaknesses
  • The technical contribution is somewhat limited. The techniques used are common: pseudo-label-based self-training, real-data replay, random exploration, and iterative updates. Please clarify the key difference from existing test-time adaptation and self-training approaches.

After Rebuttal The authors clarified that the Asymmetric adaptation is tailored to AI-generated image detection, which is sound

  • The comparisons may not be fair Most results in Table 1 compare the adapted detector with previously reported results from other methods. However, the proposed method benefits from repeated optimization on the target distribution and access to a trusted real-image set. I mean the unlabeled test set is used. It is difficult to determine how much of the improvement comes from the proposed strategy rather than the additional adaptation. Moreover, the approach assumes access to an unlabeled target collection and an independent set of trusted real images, and it repeatedly updates the detector using these data. This differs from settings where images arrive individually or where the real-image distribution changes over time.

After Rebuttal. Responses to 1) W2.1: Held-out generalization and the effect of additional real data, 2) W2.2: Comparison with TTC under a controlled adaptation budget, and 3) W2.4: Robustness of trusted-real replay to target-real distribution shift are helpful. l. Please incorporate these experiments and discussion in the revision

  • The stochastic exploration strategy needs stronger justification The method randomly selects images from the predicted-real pool and assigns them synthetic labels, even though some are genuine real images. While the results show that exploration can improve recall, it is unclear why uniform random sampling is preferable to more targeted alternatives, such as uncertainty-based selection.

After Rebuttal. Uncertainty-based Selection is good as well, which solves my question.

Questions for authors

It would be better to compare test-time adaptation approaches using the same initial detector, target data, and adaptation budget

4 comments
AuthorsOpenReview ↗

Response to Reviewer QvyL (1/3)

We sincerely thank the reviewer for the constructive comments. We have carefully addressed the concerns regarding the distinction from existing adaptation and self-training methods, comparison fairness, real-distribution shift, deployment, and exploration strategy by providing further clarifications and conducting additional controlled experiments.

Response to W1: Distinction from Existing Adaptation and Self-training

  • W1.1: Asymmetric adaptation tailored to AI-generated image detection. We design an asymmetric adaptation framework around the distinct roles of real and fake supervision in this task. Under generator shift, the predicted-real subset contains missed fake images. Conventional symmetric pseudo-labeling assigns real labels to these samples and reinforces false negatives, even after confidence filtering. Our framework separates the supervision sources, using target fake predictions for source-specific learning and trusted real supervision to counteract the false positives contained in the pseudo-fake set. As summarized below, confidence filtering cannot eliminate false negatives in the predicted-real set, so symmetric adaptation remains clearly inferior, while our asymmetric framework consistently achieves the best balanced accuracy on all three benchmarks.
Target-supervision strategy  GenImage  AIGIBench  Chameleon 
Initial detector 87.581.486.3
All predicted-real targets 89.682.088.0
High-confidence symmetric 87.584.489.3
High-confidence symmetric + replay 89.789.193.5
Ours, asymmetric 99.399.498.0
  • W1.2: Trusted supervision and stochastic exploration for difficult unseen-generator generalization. Confidence-driven adaptation selects adaptation samples according to the initial model's predictions, which can overlook difficult unseen fakes confidently classified as real. Our stochastic exploration samples beyond this confidence-based coverage, while trusted real supervision constrains the noise introduced by assigning fake labels to exploratory samples. The discovered hard fake samples then provide new source-specific anchors for subsequent iterations, progressively expanding the detector's fake-class decision region. This mechanism reduces dependence on initial confidence and source-training coverage, supporting adaptation to difficult categories that the initial detector poorly recognizes. Figure 5dd provides direct evidence on Treasure, where adding exploration increases Nano-Banana-Pro accuracy from 57.0% to 84.0%. Figure 5ee further shows faster reduction in the number of sources below 90% accuracy. Together, asymmetric supervision addresses detection-specific error reinforcement, while exploration under trusted-real constraints improves recovery of difficult unseen categories.

Response to W2 / Question: Comparison Fairness, Real-Distribution Shift, and Deployment

  • W2.1: Held-out generalization and the effect of additional real data. The paper separates each benchmark into 80% unlabeled adaptation data and 20% held-out images, which are excluded from adaptation. The held-out results show that adaptation does not fit the adaptation subset itself but learns a transferable within-class pattern of the target source that generalizes to new target images. Table 7 in Appendix E further expands the real-image sources used in initial training while keeping synthetic training data and optimization fixed, and the fake accuracy of the resulting detectors does not improve. Together with the replay-equipped symmetric controls in W1.1, this shows that the gains stem from our proposed method rather than from additional real data.

  • W2.2: Comparison with TTC under a controlled adaptation budget. TTC [1] combines adaptation with reliable pseudo-label selection, parameter anchoring, and a label-free safety check. Starting from the same SDv1.4-trained initial detector, we reproduce TTC following its original method and adaptation protocol, and compare it with our method under the same training configuration, target data, and adaptation budget. Despite its selection mechanisms, TTC's gains remain limited in these experiments.

Method  GenImage  AIGIBench  Chameleon 
Initial baseline 88.2575.3885.50
TTC, three-round reimplementation 88.1573.1386.58
Ours, same adaptation budget 97.5498.0295.92
AuthorsOpenReview ↗

Response to Reviewer QvyL (2/3)

  • W2.3: Comparison with general-purpose adaptation in generative-image detection. Starting from the same DDA-trained initial detector, we reproduce Tent [2], SAR [3], and EATA [4] under the same training configuration and target data within this comparison. These entropy-based methods yield limited and inconsistent gains over the initial detector, with degradation on several method–benchmark pairs. These results indicate that general-purpose adaptation is not sufficient for unseen-generator detection, whereas our asymmetric framework is specifically designed for generalizable AI-generated image detection and therefore achieves consistent improvements.
Method  GenImage  AIGIBench  Chameleon 
Initial baseline 94.2882.6076.40
Tent 96.4158.2350.20
SAR 94.2882.4675.10
EATA 96.4084.2077.65
Ours 99.1499.3297.87
  • W2.4: Robustness of trusted-real replay to target-real distribution shift. Using GenImage fake images as the fixed base, we vary the target real images and the replay real images to test the effect of trusted real supervision under distribution shift. These cross-source cases show that trusted real replay provides a stable anchor even when target real images deviate from the replay distribution, so identical source distributions are not required. In addition, the asymmetry of data collection further reduces the impact of target-real shift: unknown fake images cannot be obtained in advance, whereas real images can be broadly pre-collected and augmented to cover more real distributions.
Target real Replay real  Fake accuracy  Real accuracy  Balanced accuracy 
ImageNet Open Images 97.6599.5098.58
MS COCO ImageNet 98.47100.0099.24
CC12M LAION 98.1397.3897.75
Open Images CC12M 97.1398.8898.00
  • W2.5: Dual-track streaming detection and offline self-updating. Our method targets a dual-track deployment rather than real-time streaming adaptation. In the online track, any existing detector can be deployed to process incoming images, while the unlabeled images and their detection outcomes are accumulated. Once a deployment-specific threshold of samples and predicted-fake collection is reached, the offline track activates our adaptation framework to update a separate model copy, and the updated checkpoint is then returned to the online system as the new detector.

Response to W3: Random Exploration versus Uncertainty-based Selection

  • W3.1: Difficult-sample coverage and information gain across confidence regions. We partition the predicted-real confidence into five equal-width intervals, where a value near 0.5 indicates proximity to the fake decision boundary and a value near 0 indicates confident prediction as real. On the Treasure benchmark, we analyze the difficult categories with low detection accuracy, and find that most misclassified fakes of these generators fall into the high-confidence real region, largely overlapping with genuine real images rather than concentrating near the decision boundary. Uncertainty sampling therefore only covers a small and unrepresentative part of the difficult samples, and the candidates it selects lie close to already correctly classified fakes, offering limited information gain. Random exploration instead samples uniformly across confidence regions and covers a broader range of difficult samples.
Generator  0.4–0.5  0.3–0.4  0.2–0.3  0.1–0.2  0.0–0.1 
CogView4 7.276.237.2711.4267.82
GPT-image-1.5 3.924.575.557.5078.47
Seedream 4.0 4.584.915.408.5176.60
Sora-image 4.535.665.836.8077.18
Midjourney V7 3.446.076.238.3675.90
Nano-Banana-Pro 2.863.164.066.7783.16
AuthorsOpenReview ↗

Response to Reviewer QvyL (3/3)

  • W3.2: Faster convergence across benchmark settings. Building on the coverage advantage of random exploration in W3.1, we evaluate how this advantage translates into adaptation efficiency. Under the same configuration and setting, random exploration terminates earlier on all four benchmarks, and the final balanced accuracy below, averaged over multiple random seeds, shows that it still achieves better adaptation results despite the faster convergence, consistent with Figure 5 in the paper. This improvement stems from the mining of difficult samples, which substantially refines the decision boundary and thus improves recall while accelerating convergence.
Benchmark  Stopping round, random / uncertainty  Final BAcc %, random / uncertainty 
GenImage  4.00 / 8.67  99.12 / 98.02 
AIGIBench  3.67 / 4.67  99.34 / 97.97 
Chameleon  3.67 / 9.33  97.74 / 96.42 
Treasure  4.00 / 7.67  98.78 / 96.37 

[1] Yiqian Zhang et al. Test-Time Curriculum for Open-Set AIGC Detection. arXiv:2608.00559, 2026.

[2] Dequan Wang et al. Tent: Fully Test-Time Adaptation by Entropy Minimization. ICLR, 2021.

[3] Shuaicheng Niu et al. Towards Stable Test-Time Adaptation in Dynamic Wild World. ICLR, 2023.

[4] Shuaicheng Niu et al. Efficient Test-Time Model Adaptation without Forgetting. ICML, 2022.

Thank you for your constructive feedback and for revisiting your assessment. We are glad that our responses clarified the motivation for asymmetric adaptation and answered your question about uncertainty-based selection. As requested, we will incorporate the held-out evaluation, the comparison with TTC under a matched adaptation budget, and the replay robustness experiment into the revision. We appreciate your positive reassessment and the score increase.

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

This paper proposes a self-training based Asymmetry-Guided Iterative Adaptation framework for detecting AI-generated images from unseen generative models. The method treats target images initially predicted as fake as pseudo labeled supervision and iteratively adapts the detector to recover fake images from the same sources that were previously missed. To overcome the limited coverage of initial predictions on difficult sources, Stochastic Exploration provisionally labels a small random subset of the remaining target images as fake, while Trusted Real Replay supplies independent real supervision that suppresses false positives and decision boundary drift caused by noisy pseudo labels. In addition, Asymmetric Supervision withholds real pseudo labels from target images predicted as real, preventing missed fake images from being reinforced as real. As a result, the detector can adapt to an unlabeled target distribution without target annotations or generator identities. Extensive experiments on thirteen benchmarks covering GAN, diffusion, and autoregressive generators show substantial improvements over existing methods, along with robustness to common post-processing and consistent gains across diverse detector backbones.

Critical strengths
  1. The paper addresses a fundamental challenge in AI-generated image detection: keeping pace with rapidly evolving generative models. It proposes a practical adaptation framework to improve detection performance on images produced by previously unseen generators.

  2. The proposed method combines Asymmetric Supervision, Stochastic Exploration, and Trusted Real Replay to iteratively recover AI-generated images that the detector initially failed to detect. Notably, it performs adaptation without target annotations or generator identities, which constitutes a meaningful methodological contribution.

  3. The paper presents extensive experiments across 13 benchmarks covering diverse generative architectures and models, demonstrating strong detection performance that surpasses existing methods.

  4. The proposed framework is not restricted to a specific detector architecture. It consistently improves performance across various backbones, including ResNet, CLIP, DINOv2, and DINOv3, demonstrating its broad applicability and generalizability.

Critical weaknesses
  1. The training configuration of the initial detector is not clearly described. In particular, it is unclear whether the backbone is updated during source training or whether only the classification head is trained. Consequently, the reported performance gains are not sufficiently disentangled: it remains unclear whether they stem from LoRA fine-tuning of DINOv3 ViT-L itself or from the proposed iterative adaptation strategy. Controlled experiments that compare the proposed method against source-only LoRA fine-tuning and standard target self-training, under the same backbone and training configuration, would help clarify the contribution of the proposed method.

  2. The proposed method does not explicitly correct erroneous pseudo-labels arising from false positives. Instead, it mitigates their influence through Trusted Real Replay. However, the paper provides limited analysis of how such label noise affects adaptation and of the mechanism by which Trusted Real Replay counteracts it. Furthermore, the stopping criterion, which is based on the growth in the number of predicted-fake images, cannot distinguish genuine improvements in fake recall from increases in false positives.

  3. Although the paper evaluates adaptation on target distributions containing a mixture of generative models, adaptation is performed independently for each target benchmark. It therefore remains unclear whether detection performance on previously adapted generators is preserved when the detector is sequentially adapted to new generators. In particular, the possibility of catastrophic forgetting is not examined.

  4. The proposed method assumes access to a predefined unlabeled target collection, but its applicability to realistic settings in which images arrive continuously is not sufficiently discussed. In particular, the inference procedure before sufficient target data has been collected, as well as the criterion for deciding when to perform adaptation, are not clearly specified.

Questions for authors
  1. Could the authors clarify whether the DINOv3 ViT-L backbone is frozen during source training and whether both LoRA parameters and the classification head are updated during target adaptation? To isolate the contribution of the proposed strategy, could the authors compare it with source only LoRA fine-tuning and conventional LoRA-based self-training using the same target data?
  2. How do false positive pseudo labels affect iterative adaptation, and how does trusted real replay mitigate their impact? An analysis of pseudo label precision and false positive rates across iterations would be helpful. Additionally, how reliable is the stopping criterion based on predicted-fake growth, which cannot distinguish improved fake recall from increased false positives?
  3. Although mixed-source adaptation is evaluated, each benchmark is adapted independently. Can sequential adaptation to new generators preserve performance on previously adapted generators? A comparison of independent, joint, and sequential adaptation would help assess potential catastrophic forgetting.
  4. How would the proposed method operate in real-world scenarios where target images arrive continuously rather than as a predefined collection? Could the authors clarify how inference is performed before sufficient samples are available, when adaptation is triggered, and the associated computational overhead?
  5. Although the paper includes a Reproducibility Statement, code and checkpoints appear to be unavailable. Given the importance of reproducibility for assessing the reliability of the reported results and facilitating future research, do the authors plan to release the implementation code, trained checkpoints, training configurations, and details of the trusted real replay dataset?

I appreciate the authors' efforts to advance research in AI-generated image detection. I would be willing to reconsider my evaluation and potentially increase my score if the questions and concerns raised above are adequately addressed, particularly regarding the added value of the proposed adaptation strategy over standard LoRA fine-tuning and the reproducibility of the reported results.

11 comments
AuthorsOpenReview ↗

Response to Reviewer kxuB (1/2)

We sincerely thank the reviewer for the careful reading and constructive feedback. We have carefully addressed the concerns regarding the LoRA fine-tuning configuration, pseudo-label noise and stopping, sequential adaptation and forgetting, streaming deployment, and reproducibility by providing further clarifications and conducting additional experiments.

Response to W1 / Q1: LoRA fine-tuning and the contribution of adaptation

All of our DINOv3 ViT-L settings, including the source-only baseline, use the same fine-tuning configuration: the pretrained backbone is frozen, and both the LoRA adapters and the classification head are updated during source training and target adaptation. The source-only baseline is the resulting detector at round 0, and we will clarify this in Appendix A. Under this shared configuration, Table 5 of the paper already isolates the contribution of our framework: our method raises mean balanced accuracy from 85.1% to 98.9%, a gain of 13.8% over the source-only baseline. This attributes the gains to the proposed framework rather than to LoRA fine-tuning.

Balanced accuracy %  AIGIBench  Chameleon  GenImage  Avg. 
Source-only LoRA (round 0) 81.486.387.585.1
Symmetric self-training allpredicted−realtargetsall predicted-real targets 82.088.089.686.5
High-confidence symmetric self-training + trusted replay 89.193.589.790.7
Ours (asymmetric adaptation) 99.498.099.398.9

Response to W2 / Q2: False positives, trusted real replay, and stopping

  • W2.1: Counteracting false positives with trusted real replay. Results across 13 benchmarks (Tables 1–3) show that false positives have limited impact under our framework: on the 11 main benchmarks, mean fake recall rises from 79.7% to 98.4% while mean real accuracy rises from 94.7% to 98.5%, reaching 99.4% on AIGIBench and 99.6% on Treasure (Appendix F). The round-wise results in W2.3 further show that FPR generally decreases or stays low across adaptation rounds. Together, these results indicate that trusted real replay is effective in counteracting false positives.

  • W2.2: Controlled binary classification with injected false positives. To directly verify whether clean real supervision can counteract false positives, we design a controlled binary classification experiment with SDv1.4 images as fake and ImageNet images as real. We replace a fraction of the fake-class training images with ImageNet images while retaining the fake label for those replacements, simulating false positives. The results show that the increase of injected false positives does not degrade overall accuracy or FPR: even when half of the fake-labeled set consists of real images, overall accuracy remains 99.80% and FPR is 0.13%, supporting that trusted real replay can effectively counteract the effect of false positives on classification.

FP ratio %  Overall accuracy %  Fake recall %  FPR % 
0% 99.9599.900.00
5% 99.8099.630.03
10% 99.3298.730.10
20% 99.9299.830.00
50% 99.8099.730.13
  • W2.3: Validity of the growth-based stopping criterion. We supplement experiments to verify the rationality of this criterion. As shown below, FPR generally decreases or stays low across rounds, and trusted real replay constrains false positives throughout adaptation. With FPR low and stable while the predicted-fake set keeps growing, true positives dominate the growth, indicating that the growth is primarily driven by fake recall, which supports the growth-rate threshold as a reliable stopping criterion.
Benchmark  Overall accuracy % by round  FPR % by round 
GenImage  88.25 → 94.81 → 96.77 → 98.64 → 99.24  0.29 → 0.08 → 0.08 → 0.25 → 0.19 
AIGIBench  75.38 → 88.14 → 97.14 → 97.67 → 97.98  3.50 → 1.17 → 1.12 → 0.83 → 0.50 
Chameleon  85.50 → 94.00 → 96.17 → 98.54 → 99.16  2.67 → 1.83 → 1.17 → 0.67 → 0.25 
AuthorsOpenReview ↗

Response to Reviewer kxuB (2/2)

Response to W3 / Q3: Sequential adaptation and forgetting

We compare independent, joint, and sequential adaptation across eight GenImage sources in forward and reverse orders. Sequential adaptation shows only modest forgetting ∗∗960˙7**96\.07%/96\.57%** vs\. **98\.61%/98\.33%** for independent and **99\.19%/99\.21%** for joint, and simply replaying fake examples from earlier stages restores accuracy to 98.61%/98.29%, comparable to joint adaptation in both orders. This is consistent with classical continual-learning findings that a replay buffer effectively counteracts forgetting.

Setting  Forward overall accuracy %  Reverse overall accuracy % 
Independent source−specificmodelssource-specific models 98.6198.33
Joint 99.1999.21
Sequential 96.0796.57
Sequential + historical fake replay 98.6198.29

Response to W4 / Q4: Streaming deployment and adaptation triggering

  • W4.1: Dual-track deployment. Our method targets a dual-track deployment rather than real-time streaming adaptation. Before adaptation is triggered, incoming images can be detected by any existing detector deployed online, and this detector serves as the starting point from which our framework progressively improves detection ability through iterative adaptation. The online system detects incoming images in a streaming manner while accumulating images and their detection outcomes. Once the accumulated unlabeled data reaches a deployment-specific threshold, the offline adaptation framework is activated, and the updated checkpoint is pushed back to the online system.

  • W4.2: Threshold as a deployment trade-off. The collection-size threshold trades update frequency against coverage of unknown generators. A smaller threshold enables faster adaptation to emerging generators, whereas a larger threshold accumulates more diverse evidence for stronger adaptation. Appendix D characterizes the sensitivity to target-set size and composition and provides guidance on when adaptation is preferable and can yield reliable gains.

  • W4.3: Open and detector-agnostic framework. Our framework does not restrict the detector architecture or detection pipeline. It provides a dynamic self-update mechanism: any static online detector can be iteratively updated offline by our framework, improving its detection ability on unseen data. The architecture ablation in Appendix C supports this detector-agnostic property.

  • W4.4: Computational overhead. Since adaptation is performed offline, it does not affect the online streaming detection path. The offline update cost is therefore acceptable in practice and does not compromise online detection latency.

Response to Q5: Reproducibility and release

We thank the reviewer for raising this point and fully agree on the importance of reproducibility. We will release the implementation, source-trained and adapted checkpoints, training configurations, target splits and seeds, and the details of the trusted real replay pools (composition, construction procedure, and deduplication) upon acceptance.

Thank you for the thoughtful responses and additional experiments. Some of my concerns have been addressed, but I would appreciate further clarification on the following points.

W1 / Q1: Training of the initial detector and the contribution of adaptation

  1. In Table 13 of DDA, DINOv2 ViT-L + LoRA achieves 82.4%/91.7% on Chameleon/GenImage. In comparison, this paper’s initial detector, DINOv3 ViT-L + LoRA, achieves 87.5% on GenImage, 4.2%p below DDA. Furthermore, Table 6 reports 58.8%/76.9% for DINOv2 ViT-L on Chameleon/GenImage. Could you explain whether these differences arise from the training data and alignment procedures, LoRA settings, input preprocessing, or evaluation protocols? I would like to understand whether the initial detector was sufficiently trained and how strongly the reported adaptation gains depend on its initial performance.
  2. To help assess whether the initial DINOv3 ViT-L detector was sufficiently trained, could you provide its detailed training configuration? In particular, please specify the training epochs/iterations, optimizer, learning rate and scheduler, batch size, weight decay, LoRA settings, validation split, checkpoint selection criterion, and whether early stopping was used. This information would help clarify the initial training conditions and the dependence of the reported adaptation gains on the initial detector’s performance.

W2.2: Controlled experiment with injected false positives

Was this experiment initialized from the source-trained detector? If so, could you report its performance on the same SDv1.4/ImageNet evaluation set before any updates? If the detector has already learned traces associated with SD VAEs, SDv1.4 may be relatively easy to distinguish, making it difficult to separate the robustness of its existing representations from the effect of Trusted Real Replay. B-Free [1] and AlignedForensics [2] are also relevant prior studies in this regard.

It would therefore be helpful to compare performance with and without replay under the same initialization, proportion of injected false positives, and training budget. Please also specify the sample counts, clean-to-noisy data ratio, loss weighting, and number of updates. If feasible, please examine whether the effect also holds for difficult unseen generators or real images that the detector actually misclassifies as fake.

Q5: Reproducibility and release

Thank you for sharing your release plans. To enable an assessment of reproducibility during the current review process, could you provide the following essential information?

  1. Initial training configuration: epochs/iterations, optimizer, learning rate and scheduler, batch size, weight decay, validation split, checkpoint selection criterion, and early stopping.
  2. LoRA settings: rank, scaling factor, and target layers/modules.
  3. Backbone specifications: the specific DINOv3 ViT-L model variant and configuration, including patch size/stride and whether register tokens are used.
  4. Replay pool composition: the sources of the datasets and web photographs, image counts per source, and total pool size. If full disclosure is not possible, please provide details to the extent that they can be shared.

Please correct me if I have misunderstood any experimental settings. I will promptly review and respond to your follow-up comments and continue engaging in the discussion before finalizing my assessment.

[1] A Bias-Free Training Paradigm for More General AI-generated Image Detection. (CVPR 2025)

[2] Aligned Datasets Improve Detection of Latent Diffusion-Generated Images. (ICLR 2025)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.