💻 computer science

Test-Time Curriculum for Open-Set AIGC Detection

This paper proposes Test-Time Curriculum (TTC), a model-agnostic framework that enhances open-set AIGC detection under unseen generator shifts by adaptively refining pseudo-labels through a curriculum-based self-training process and cross-scale evidence aggregation, validated on a new comprehensive benchmark called AIGCGuard.

Original authors: Yiqian Zhang, Zheyuan Gu, Xiangzhao Hao, Zefeng Zhang, Jingjia Mao, Jiahao Hu, Jiaxu Miao, Jun Yu, Zhenyu Zhang, Shuohuan Wang, Yu Sun

Published 2026-08-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yiqian Zhang, Zheyuan Gu, Xiangzhao Hao, Zefeng Zhang, Jingjia Mao, Jiahao Hu, Jiaxu Miao, Jun Yu, Zhenyu Zhang, Shuohuan Wang, Yu Sun

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a security guard at a massive, ever-changing art museum. Your job is simple: spot the forgeries. But here's the twist—the forgers aren't just one guy with a bad brush; they are a legion of AI robots that upgrade their skills every single day. Yesterday's fake paintings looked a bit blurry, but today's are so perfect they could fool a human expert. The problem is, you were trained only on the old, blurry forgeries. When a new, super-smart robot forger walks in, your old training doesn't work, and you start letting fakes slip right past you. This is the daily struggle of "AIGC detection"—the science of spotting AI-generated images. For a long time, scientists tried to build a "super-guard" that could spot any fake from the start. But the paper you're about to read suggests that's the wrong approach. Instead of trying to be a genius who knows everything beforehand, what if your guard could learn on the job? What if, the moment a new forger arrives, the guard could quickly study the new fake art, figure out the new tricks, and update their own rules without needing a human teacher to hand them a new textbook?

This is exactly what the researchers behind this paper, titled "Test-Time Curriculum for Open-Set AIGC Detection," have figured out. They propose a clever new system called TTC (Test-Time Curriculum). Think of TTC not as a rigid rulebook, but as a smart, self-teaching tutor that helps a detector adapt to new AI generators the moment they appear. The core idea is that instead of trying to learn from every new image at once (which is risky because the detector might get confused by bad guesses), the system starts with the easiest, most obvious cases. It learns from those, gains confidence, and then slowly moves on to the trickier, harder-to-spot fakes. It's like a video game where you don't jump straight to the final boss; you beat the easy levels first to get your gear ready, then tackle the hard ones.

The paper introduces a specific method to make this learning safe and effective. First, it uses a "Cross-Scale Pseudo-Label Refinement." Imagine looking at a painting: sometimes you can spot a fake by zooming in on a tiny brushstroke, and other times by stepping back to see the whole composition. TTC looks at the image at multiple sizes at once, combining these clues to make a much smarter guess about whether an image is real or fake. Then, it uses a "curriculum" approach. It picks out the images the detector is most sure about (the "reliable" ones) and uses those to teach the detector. Crucially, it balances this learning so the detector doesn't just learn to spot fakes and forget how to spot real photos, or vice versa. As the detector gets better, the curriculum gets harder, feeding it images that are just on the edge of being confusing, helping it sharpen its skills.

To prove this works, the team didn't just test it on old data; they built a brand-new, massive playground called AIGCGuard. This benchmark contains 3,100 real images and a whopping 124,000 images generated by 40 of the most advanced AI models currently in existence, including both open-source and secret, proprietary ones. When they tested their TTC method against existing detectors on this tough new ground, the results were impressive. On average, their method boosted detection accuracy by a significant margin, often turning detectors that were failing miserably on new AI models into champions. For instance, on one challenging benchmark, their method pushed accuracy from a shaky 82.9% up to a near-perfect 99.1%.

The paper suggests that this "learning on the fly" approach is a game-changer. It argues that the old way—training a model once and hoping it lasts forever—is doomed to fail as AI generators keep getting smarter. By using this curriculum-based self-training, detectors can stay one step ahead of the evolving forgers without needing a human to retrain them from scratch every time a new AI model drops. The authors show that this method is robust, working well even when images are compressed or resized, and it works across different types of detector "brains." While the paper notes that this is currently an "offline" process (meaning it adapts to a batch of images all at once rather than streaming in real-time), it lays the groundwork for a future where our digital security guards can learn, adapt, and stay sharp in a world where the rules of reality are constantly being rewritten by machines.

Technical Summary: Test-Time Curriculum for Open-Set AIGC Detection

Problem Statement

The rapid advancement of AI-generated image (AIGC) models has created a critical challenge for visual data ecosystems: the "open-set generator shift." Existing detectors are typically trained on a limited set of known generators and deployed as static models. When new, stronger, or different generative models emerge, the test distribution shifts, causing previously learned artifacts to disappear or change. Consequently, static detectors often fail to generalize to unseen generators. While existing methods improve generalization through better representations or training data construction, they adhere to a "train-once-and-deploy" paradigm and cannot adapt to emerging distributions without costly retraining or labeled target data.

Methodology

The authors propose Test-Time Curriculum (TTC), a model-agnostic framework that enables detectors to adapt to unlabeled test data from unseen generators using a curriculum-based self-training approach. The methodology consists of two primary components:

1. Test-Time Curriculum Adaptation Framework

Instead of adapting to all test data at once, TTC employs a multi-round self-training process that progresses from highly reliable samples to more challenging, informative cases.

Reliable Selection: The framework selects pseudo-labeled samples based on prediction entropy. Samples with low entropy (high confidence) are considered reliable.
Class-Balanced Selection: To prevent gradient bias toward the class that is easier to classify (often real images), TTC enforces an equal number of selected samples from both the "reliable fake" and "reliable real" classes.
Hard-Within-Reliable Sampling: Within the balanced reliable set, the framework prioritizes samples with higher uncertainty (closer to the decision boundary) to maximize the informativeness of the adaptation.
Curriculum Progression: The adaptation proceeds in rounds. Early rounds use strict confidence thresholds to ensure safety. Subsequent rounds relax these thresholds, allowing the detector to incorporate harder samples as it aligns closer to the target distribution.
Parameter Anchoring: The loss function includes an $L_2$ regularization term to constrain the detector's parameters from drifting too far from the initial pre-trained weights, preventing catastrophic forgetting or noise-induced collapse.

2. Cross-Scale Pseudo-Label Refinement

To address the unreliability of single-scale predictions, TTC integrates a multi-scale prediction strategy:

Multi-Scale Prediction: Images are processed at multiple resolutions (full image, downsampled, and local patches). Patch-level predictions are aggregated using logit-space log-sum-exp pooling to form scale-level probabilities.
Cross-Scale Fusion for Pseudo-Labels: For selecting reliable samples and assigning pseudo-labels, the framework averages predictions across all scales. This consensus reduces noise caused by calibration inconsistencies or specific generator artifacts at a single resolution.
Noisy-Or Fusion for Inference: During final inference, the framework combines scale-specific predictions using a noisy-or fusion rule. This heuristic emphasizes images that receive a strong "fake" signal at any scale, strengthening the final prediction without requiring calibrated probabilities.

Key Contributions

Formulation of Open-Set AIGC Detection as TTA: The paper reframes open-set AIGC detection as a test-time adaptation (TTA) problem, enabling detectors to adjust to unlabeled data from unseen generators without additional annotation or retraining from scratch.
Test-Time Curriculum (TTC) Framework: A novel, multi-round self-training framework that combines class-balanced reliable selection, hard-within-reliable sampling, parameter anchoring, and cross-scale consensus. It operates without modifying the underlying detector architecture.
AIGCGuard Benchmark: The construction of a new benchmark containing 3,100 representative real images and 124,000 generated images from 40 state-of-the-art open-source and proprietary text-to-image models. This dataset covers diverse semantic categories, resolutions, and generator families to support evaluation under realistic generator shifts.

Experimental Results

Extensive experiments were conducted on five benchmarks: AIGCDetectionBenchmark, AIGI-Holmes, GenImage, Chameleon (in-the-wild), and the proposed AIGCGuard.

Performance Gains: TTC consistently outperformed existing baselines, including strong data-alignment methods like DDA. On the AIGCDetectionBenchmark, TTC achieved a balanced accuracy (BAcc) of 93.0%, surpassing the Base detector by 11.4 points and DDA by 5.2 points.
Robustness to Unseen Generators: Significant improvements were observed on challenging generators such as Midjourney (+32.2% over Base), StarGAN (+27.2%), and VQDM (+26.9%). On the AIGI-Holmes benchmark, TTC reached 99.1% BAcc, effectively handling diverse generator families including autoregressive and diffusion models.
Ablation Studies: The study confirmed that the components of TTC are complementary. Balanced Reliable Selection (BRS) provided the initial gain, while Cross-Scale Pseudo-Label Refinement (CSPLR) was crucial for enabling effective "hard-within-reliable" sampling.
Robustness: TTC maintained strong performance under image perturbations (JPEG compression, resizing, Gaussian blur) and demonstrated stability across multiple adaptation runs.

Significance and Claims

The paper claims that Test-Time Curriculum (TTC) establishes a practical and effective framework for maintaining AIGC detector robustness in rapidly evolving real-world environments. By shifting the focus from static cross-generator generalization to adaptation on unlabeled deployment data, TTC addresses the core difficulty of open-set generator shift without the prohibitive costs of continuous data collection and retraining.

The authors emphasize that their approach is particularly attractive for industrial applications where data curation pipelines periodically process unlabeled image collections reflecting recent generator distributions. Leveraging these data at test time reduces the need for costly cycles of generator-specific annotation. The introduction of AIGCGuard further supports the community by providing a modern, high-fidelity benchmark that better reflects the current state of text-to-image generation, facilitating more realistic evaluation of open-set detection capabilities.

The work concludes that test-time adaptation is a viable direction for open-set AIGC detection, provided that adaptation is conducted via a curriculum that carefully manages pseudo-label reliability and class balance to prevent error propagation.