acceptodds
Withdrawn from ICLR 2027

Test-Time Curriculum for Open-Set AIGC Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

AI-generated image detectors deployed in open-world environments inevitably face distribution shifts as new and stronger generative models continue to emerge. Although existing methods improve cross-generator generalization through better representations or training data construction, they typically follow a static train-once-and-deploy paradigm and cannot adapt after deployment. In this work, we study open-set AIGC image detection from a test-time adaptation perspective, where pretrained detectors adapt to unlabeled target data from unseen generators. We propose Test-Time Curriculum (TTC), a simple and model-agnostic framework that performs curriculum-based self-training for open-set AIGC detection. TTC separates pseudo-label reliability from adaptation difficulty by starting from reliable balanced samples and progressively incorporating harder informative cases under generator shift. We further introduce Cross-Scale Pseudo-Label Refinement to improve pseudo-label reliability by aggregating complementary evidence across multiple resolutions. In addition, we construct AIGCGuard, a benchmark containing 3,100 representative real images and 124,000 generated images from 40 of the most advanced open-source and proprietary text-to-image models. Extensive experiments on five benchmarks demonstrate that TTC substantially improves detection performance under diverse unseen-generator shifts.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
AIGC Detection, Test-Time Adaptation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_26cbOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the open-set generator shift in reliable AIGC detection. It proposes test-time curriculum (TTC) to perform curriculum-based self-training and pseudo-label refinement by aggregating evidence across resolutions. It also constructs a benchmark for AIGC detection.

Critical strengths

1.The paper is well-written and easy to follow.

2.The paper studies an interesting issue and proposes a new AIGC detection benchmark. Experimental results shows that the proposed method is effective across different benchmarks.

Critical weaknesses

1.Using binary entropy to quantify reliability is questionable because in TTA, a prediction farther from the decision boundary is not necessarily trustworthy. The model could give confidently wrong predictions. The selection of the uncertainty threshold \tau_r also needs systematic analysis.

2.The main framework of TTC is based on curriculum learning. However, the paper lacks a detailed discussion of curriculum learning.

Questions for authors

Could you explain how does your selection criteria (binary entropy) address confidently wrong predictions?

Reviewer_fiVDOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Test-Time Curriculum (TTC), a test-time adaptation framework for detecting AI-generated images from unseen generators. TTC combines balanced reliable sample selection, hard-within-reliable curriculum learning, and cross-scale pseudo-label refinement to improve adaptation under distribution shifts. The authors also introduce AIGCGuard, a benchmark containing 127,100 images from 40 recent generative models. Experiments on five benchmarks demonstrate significant improvements over existing AIGC detection methods, supported by ablation studies and robustness evaluations.

Critical strengths
  1. The paper addresses the important challenge of adapting AI-generated image detectors to unseen generators after deployment. Introducing test-time adaptation into open-set AIGC detection is practically valuable, especially as generative models continue to evolve.

  2. TTC achieves substantial improvements across multiple benchmarks, including AIGCDetection, AIGI-Holmes, and AIGCGuard. The ablation studies and robustness experiments further support the effectiveness of the proposed components.

Critical weaknesses
  1. The main components of TTC, including confidence-based pseudo-label selection, class balancing, curriculum self-training, and multi-scale prediction aggregation, are largely based on established techniques. Although their combination achieves promising results, the methodological novelty appears limited.

  2. The main experiments use the entire unlabeled test set for adaptation and evaluate on the same samples, raising concerns about generalization to unseen data and practical deployment. Although the authors provide an additional non-overlapping evaluation, its scope is limited. Furthermore, the source code and the proposed AIGCGuard benchmark have not been released with the submission, making it difficult to independently verify and reproduce the reported results.

Questions for authors

Q1. How does TTC fundamentally differ from existing confidence-based curriculum learning and pseudo-label self-training methods? Could the authors further clarify the methodological novelty?

Q2. Since the main evaluation adapts and tests on the same dataset, could the authors provide more comprehensive results using strictly separate adaptation and evaluation sets, particularly across different unseen generators?

Q3. The source code and AIGCGuard benchmark are not currently available. Could the authors provide anonymized code and sufficient benchmark construction details to support independent reproduction of the reported results?

Reviewer_Gc6XOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies test-time adaptation for AI-generated image detection under unseen-generator shifts. The authors propose Test-Time Curriculum (TTC), which progressively adapts a pretrained detector using unlabeled target data. The method combines class-balanced reliable pseudo-label selection, hard-within-reliable sample selection, and cross-scale pseudo-label refinement. The paper also introduces AIGCGuard, a benchmark covering 40 recent image generators. Experiments across five benchmarks demonstrate substantial improvements over the initial detector and several existing detection methods.

Critical strengths
  1. The paper studies a relevant problem. Adapting pretrained detectors to emerging generators without labeled target data is an interesting direction, and the proposed framework is relatively simple and applicable to different detector architectures.

  2. The experimental results are strong across multiple benchmarks.

Critical weaknesses
  1. The primary evaluation setting is transductive, which limits the conclusions about adaptation to evolving generators. The main experiments allow TTC to adapt using the full unlabeled target test set and then evaluate on the same collection. This is a valid transductive setting, but it differs from the deployment scenario emphasized in the paper, where new generators emerge over time.

  2. The effectiveness of hard-within-reliable selection depends substantially on the quality of the pseudo labels. Table 3 shows that adding hard-within-reliable selection to balanced reliable selection reduces balanced accuracy from 92.9% to 83.8%. Performance improves only after incorporating cross-scale pseudo-label refinement. While the full-model ablation supports the usefulness of HWR when combined with CSPLR, the large performance degradation without refinement raises questions about the robustness of the curriculum strategy.

  3. The contribution of parameter adaptation should be more clearly separated from multi-scale inference. The proposed framework combines parameter adaptation with multi-scale prediction and noisy-or fusion at inference. These components can affect classification performance independently.

  4. The hard-within-reliable selection rule deserves further clarification. According to Eqs. (4)–(5), the number of selected samples per class is determined by the size of the smaller reliable class. Consequently, all reliable samples from the minority predicted class are retained, while hardness-based selection is effectively applied only to the larger class.

  5. The comparison with recent methods could be strengthened. Most of the AIGC detection baselines considered in the paper are from 2023–2025, while several more recent methods are not included. Although the appendix provides comparisons with representative test-time adaptation methods, these are primarily general-purpose approaches rather than methods specifically developed for generated-content detection.

Questions for authors
  1. How does the pseudo-label accuracy of HWR-selected samples compare with random or confidence-based selection within the reliable set? Can this explain the performance drop observed in Table 3?

  2. What is the performance of the frozen Base detector when the same multi-scale prediction and noisy-or fusion strategy is used, without parameter adaptation?

  3. How sensitive is HWR to the imbalance between predicted real and fake samples? Would selecting a fixed proportion of hard samples from each class provide more consistent behavior?

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

Related papers

Discussion (0)

Sign in to comment.