Test-Time Curriculum for Open-Set AIGC Detection

2026-08-01

AI Summary

This work addresses the challenge of test-time distribution shift in open-world AI-generated image detection, where emerging generative models continually degrade detector performance. To this end, the authors propose TTC, a model-agnostic test-time curriculum learning framework that enhances robustness through curriculum-based self-training on unlabeled test data—progressively expanding from high-confidence to harder samples. TTC integrates class-balanced sampling, cross-scale pseudo-label refinement, and a noise-aware probability fusion strategy to improve adaptation. As the first approach to introduce test-time adaptation into open-set AIGC detection, TTC significantly outperforms existing methods across five benchmarks. The study also introduces AIGCGuard, a new benchmark comprising 40 state-of-the-art text-to-image generators, which further validates the method’s effectiveness and generalization capability.

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. We propose Test-Time Curriculum (TTC), a simple and model-agnostic framework that adapts a detector on unlabeled test data through curriculum-based self-training. TTC starts from highly reliable pseudo-labeled samples and progressively incorporates harder yet informative cases, while enforcing class-balanced selection to reduce biased updates under generator shift. To further improve pseudo-label quality, we introduce Cross-Scale Pseudo-Label Refinement, which aggregates complementary evidence across multiple resolutions for more reliable adaptation, and applies noisy-or fusion at inference to strengthen final predictions. In addition, we construct AIGCGuard, a new 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 show that TTC substantially improves overall detection performance under diverse unseen-generator shifts, establishing a practical and effective test-time adaptation framework for open-set generated image detection.

Problem

• open-set detection
• distribution shift
• test-time adaptation
• AIGC detection
• generalization

Innovation

• Test-Time Adaptation
• Curriculum Learning
• Open-Set Detection
• Pseudo-Label Refinement
• AIGC Detection