Fleet: Few Shots Lead Effective AI-generated Image Detection

2026-06-29

AI Summary

Current methods for detecting AI-generated images rely on static feature spaces, which struggle to generalize to rapidly evolving generative models. This work proposes a dynamic few-shot adaptation framework that moves beyond the conventional static zero-shot assumption. By introducing a constrained routing correction mechanism within a disentangled subspace, the approach steers new samples away from non-AI-dominated pathways, enabling efficient online adaptation to unseen generators. With only ten reference samples, the method achieves a substantial improvement in detection accuracy—rising from 20.4% to 73.1%—on advanced generators such as Doubao Seedream 4.0, significantly outperforming existing state-of-the-art static approaches.

Abstract

AI-generated image (AIGI) detection is undergoing a critical transition from laboratory benchmarks to open-world adversarial defense. The prevalent paradigm focuses on finding static feature spaces, assuming that some invariant artifacts learned from historical data can achieve universal zero-shot generalization. While achieving saturation on several AIGI benchmarks, this static hypothesis suffers a severe performance drop against rapidly evolving generators (e.g., SD3, Nano Banana Pro). To address these limitations, we propose that the field should expand beyond "static generalization" to a new paradigm of "dynamic adaptation". We introduce Fleet, a framework that pioneers a dynamic paradigm of continuous few-shot evolution, enabling rapid alignment with emerging generative threats. Fleet improves few-shot adaptation by replacing unconstrained feature updates with constrained routing correction, where avoidance routing redirects novel AI samples away from Non-AI-dominated routes within decoupled subspaces. To validate this, we present Treasure, a benchmark spanning 64 models and 360k images, featuring diverse architectures and 20 closed-source commercial engines. Experiments reveal that while static SOTA methods fail catastrophically on modern generators, Fleet restores performance from 20.4% to 73.1% with only 10-shot adaptation on "Doubao Seedream 4.0". Code and data are available at https://github.com/ICTMCG/Fleet .

Problem

• AI-generated image detection
• static generalization
• dynamic adaptation
• few-shot learning
• adversarial defense

Innovation

• dynamic adaptation
• few-shot learning
• AI-generated image detection
• constrained routing correction
• open-world generalization