acceptodds
Under review as a conference paper at ICLR 2027

Meta-Learning for AI-Generated Image Detection

Abstract

Detectors of AI-generated images are usually trained once and then deployed unchanged.Their accuracy drops on generators that appear after training. We recast detection as a meta-learning problem. The detector conditions every prediction on a small labelled support set from the current deployment context instead of fixing a decision boundary at training time. We give a probabilistic account of how training episodes are built. The standard explicit construction keeps one class for real images and one class per generator. It needs every generator to be known and labelled in advance, as if given by an oracle. It also never pairs a query with support images of another generator, so it does not reward generalisation to a new generator. We propose an implicit construction over the binary real and AI-generated label space. Episode diversity then comes from a prior over the class-conditional distributions rather than from the choice of classes. This prior is stated apart from the data, and we give two instances of it. An empirical prior draws the class-conditional distributions from the domains of the training data. A synthetic prior draws them from Gaussian mixtures in embedding space and needs no images at all. We cast twelve meta-learning methods into this framework. With a nearest-centroid rule, they give thirteen context-aware detectors, which we compare with eleven static detectors. Our implicit construction degrades least when the support set comes from generators other than the query's. In this case, it matches or improves on the oracle explicit construction. Conditioning on the support set matters most on in-the-wild images, where static detectors lose their accuracy. A detector trained on the synthetic prior alone detects these images better than a zero-shot static detector.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
AI-generated image detection, deepfake image detection, meta-learning, in-context learning, episode construction, cross-generator generalization, synthetic priors

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_MgDMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This manuscript proposes treating AI-generated image detection as a context-conditioned meta-learning task. Instead of training a static classifier, the detector adapts its decision boundary based on a small labeled support set from the deployment environment. The authors critique the conventional explicit episode construction—where each generator forms a separate class—as requiring oracle knowledge of all generators and being unable to handle novel ones. They introduce an implicit construction that maintains binary real/fake labels while introducing diversity through priors over class-conditional distributions, instantiated via empirical training domains or synthetic Gaussian mixtures in embedding space. Thirteen context-aware detectors (twelve meta-learning methods plus nearest-centroid) are evaluated against eleven static baselines on AI-GenBench and Chameleon, with analysis of support-set utilization and different support-query relationships.

Critical strengths
  1. The paper makes a compelling case that static detection is fundamentally misaligned with the real-world deployment scenario, where new generators continuously emerge.

2.The three support policies systematically probe different degrees of support-query alignment, from ideal to realistic to challenging. This design reveals when context-conditioning helps versus when it fails, providing more nuanced insights than a simple accuracy comparison.

Critical weaknesses

1.The central claim—that the implicit construction generalizes better to unseen generators than the explicit one—is not convincingly demonstrated. The comparison confounds multiple factors: different label spaces, different support-set compositions and different aggregation strategies.

2.The paper does not adequately address a simpler alternative: taking a strong static detector and performing lightweight adaptation (e.g., linear probing or prototype calibration) on the support set at test time. The nearest-centroid baseline already achieves competitive performance, suggesting that much of the benefit comes from having labeled deployment examples rather than from the meta-learning machinery.

Questions for authors

1.Suppose a practitioner has a strong static detector (e.g., a ResNet trained on diverse generators) and 20 labeled support images from a new deployment context. What is the performance gain from: (a) using your meta-learned detector with the support set, versus (b) fine-tuning the static detector's final layer on the same 20 images? This comparison would clarify whether meta-learning is necessary or whether simple adaptation suffices.

2.The synthetic prior is an intriguing concept, but its practical utility is unclear. In what real-world scenario would a user have no training images yet need a detector? If some images are always available, does the synthetic prior offer any advantage over the empirical one, or is it primarily of theoretical interest?

Reviewer_58fnOpenReview ↗
Rating
3(+1)
Confidence
3(−1)
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents an alternative approach to conventional cross-generator generalization in AI-generated image detection by reformulating the task as a meta-learning problem. Instead of relying on a fixed classifier, the proposed framework dynamically adapts its decision boundary using a small set of labeled samples from the target deployment environment. Specifically, the authors introduce an implicit episode construction based on binary real/fake labels, where task diversity is achieved by varying class-conditional distributions. The framework considers both an empirical prior derived from training data and a synthetic prior based on Gaussian mixtures in feature space. Experimental results demonstrate that the proposed approach achieves better generalization when the support and query sets contain images from different generators.

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. The proposed method requires labeled real/fake samples from the target environment to construct the support set at test time. However, obtaining reliable labels in open-world scenarios is itself challenging, which significantly limits the practical applicability of the proposed framework.

  2. Several existing studies have already explored meta-learning for AI-generated image detection (e.g., [1]). Therefore, the authors should more clearly highlight the core contributions of this work beyond simply reformulating AI-generated image detection as a meta-learning problem.

  3. Since the primary task is AI-generated image detection, the experimental evaluation should place greater emphasis on comparisons with representative detection methods, rather than mainly examining the transfer of existing meta-learning algorithms to this task. Furthermore, the comparisons provide insufficient coverage of recent advances from 2025 and 2026.

  4. Many important experiments and analyses are relegated to the appendix, leaving several key questions unresolved after reading the main paper alone. This presentation makes it difficult to fully assess the proposed method and does not meet the expected standard of clarity for a top-tier conference paper. The authors should consider moving the most important experimental results and analyses into the main text.

[1] Xu Y, Li B, Li W, et al. Alden: Dual-level disentanglement with meta-learning for generalizable audio deepfake detection[C]//Proceedings of the 33rd ACM International Conference on Multimedia. 2025: 7277-7286.

Questions for authors

Please refer to the Weaknesses section above for detailed questions and concerns, particularly regarding the practical limitations of test-time labeled support sets, the novelty relative to existing meta-learning approaches, the completeness of comparisons with recent AI-generated image detection methods, and the presentation of key experimental results.

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

To address the problem that the accuracy of AI-generated image detectors drops when facing new generators, this paper formulates AI-generated image detection as a meta-learning framework: at prediction time, the detector conditions on a small labelled support set from the deployment context, and the image features come from a frozen CLIP ViT-L/14. The authors point out that the standard "explicit construction" (one class for real images and one class per generator) requires the labels of all generators to be known in advance, and propose an "implicit construction" that builds episodes only over the binary real/AI-generated label space. The experiments cast 12 meta-learning methods and a nearest-centroid rule into this framework and compare them with 11 static detectors.

Critical strengths
  1. The formalization in this paper is clear and useful. It naturally leads to a binary construction that requires only real/AI-generated labels and that users can actually deploy.
  2. The empirical study is broad in scope: they compare multiple detectors and design diagnostic experiments such as label inversion, label noise, pseudo-labelled support sets, and gain decomposition.
Critical weaknesses

Chameleon is evaluated only under the matched policy, where the support set and the queries are drawn from two halves of the same test set. In this setting, the nearest-centroid rule, which has no trainable parameters, reaches 0.85, only 0.01 lower than the best detector, TabPFN (0.86), and higher than all other meta-learners. Meta-training on AI-GenBench thus brings no gain beyond directly computing class means on CLIP features, and the authors should analyze this.

Questions for authors
  1. Could the authors construct an in-the-wild setting in which the support set does not come from the query distribution, e.g., with the support set from AI-GenBench and the queries from Chameleon, or with the support set from one in-the-wild source and the queries from another? This is exactly the deployment scenario in the paper's motivation, and it has not yet been tested on in-the-wild data.
  2. Why do the meta-learners underperform NearestCentroid on in-the-wild data? On Chameleon, all 11 meta-learners except TabPFN fall below NearestCentroid; the role of the components other than the feature-extraction backbone deserves discussion.
  3. If the static detectors also used the same support set to recalibrate their thresholds or logit biases, or if kNN or linear classification were performed directly on CLIP features with the same support set, how would they compare with the context-aware detectors?

Comments on the paper

1 comment
Area_Chair_adjdOpenReview ↗

A Gentle Reminder: About One Day Left in the Discussion Period

Dear Reviewers,

This is a gentle reminder that approximately one day remains in the discussion period. Please review the authors' responses and other reviewers' comments, and engage in the discussion to clarify which concerns have been addressed and which remain unresolved.

Thank you for your time and effort.

Best,
AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.