Simplicity Prevails: The Emergence of Generalizable AIGI Detection in Visual Foundation Models

TL;DR

The paper addresses the generalization gap in AIGI detection by showing that a simple linear classifier on frozen Vision Foundation Model features can achieve state-of-the-art results across standard benchmarks, unseen generators, and in-the-wild data. The method reveals two emergent mechanisms—semantic conceptualization in Vision-Language Models and data-driven feature discrimination in Self-Supervised Learning—driven by large-scale pretraining on synthetic content. Extensive experiments demonstrate strong performance, including near-perfect transfer to novel generators, challenging the dominance of specialized detectors and suggesting a paradigm shift toward utilizing foundation-model world knowledge for robust forensics. Despite these gains, the work also highlights limitations, such as blindness to VAE reconstruction and localized editing, underscoring the need for future research to balance broad generalization with fine-grained detection in real-world contexts.

Abstract

While specialized detectors for AI-Generated Images (AIGI) achieve near-perfect accuracy on curated benchmarks, they suffer from a dramatic performance collapse in realistic, in-the-wild scenarios. In this work, we demonstrate that simplicity prevails over complex architectural designs. A simple linear classifier trained on the frozen features of modern Vision Foundation Models , including Perception Encoder, MetaCLIP 2, and DINOv3, establishes a new state-of-the-art. Through a comprehensive evaluation spanning traditional benchmarks, unseen generators, and challenging in-the-wild distributions, we show that this baseline not only matches specialized detectors on standard benchmarks but also decisively outperforms them on in-the-wild datasets, boosting accuracy by striking margins of over 30\%. We posit that this superior capability is an emergent property driven by the massive scale of pre-training data containing synthetic content. We trace the source of this capability to two distinct manifestations of data exposure: Vision-Language Models internalize an explicit semantic concept of forgery, while Self-Supervised Learning models implicitly acquire discriminative forensic features from the pretraining data. However, we also reveal persistent limitations: these models suffer from performance degradation under recapture and transmission, remain blind to VAE reconstruction and localized editing. We conclude by advocating for a paradigm shift in AI forensics, moving from overfitting on static benchmarks to harnessing the evolving world knowledge of foundation models for real-world reliability.

Figures (1)

• Figure 1: The Surge of Generative Data in Web Corpora. We track the number of indexed pages from civitai.com within Common Crawl snapshots from 2022 to 2025.