Orthogonal Subspace Decomposition for Generalizable AI-Generated Image Detection

Zhiyuan Yan, Jiangming Wang, Peng Jin, Ke-Yue Zhang, Chengchun Liu, Shen Chen, Taiping Yao, Shouhong Ding, Baoyuan Wu, Li Yuan

International Conference on Machine Learning 2025 · Oral

The proliferation of sophisticated AI models capable of generating highly realistic images has introduced a critical challenge: the reliable detection of **AI-generated images (AIGI)**. This talk, presented by Zhiyuan Yan from Shanghai AI Lab on behalf of a collaborative team from Tencent Youtu Lab and Peking University, delves into the inherent difficulties of AIGI detection and proposes a novel solution: **Orthogonal Subspace Decomposition (OSD)**. The core problem lies in the "asymmetry" of the AIGI detection task, where conventional models tend to quickly overfit to specific fake patterns, leading to poor generalization to unseen or novel AI-generated content.

AI review

This paper tackles a real and important problem — generalizable detection of AI-generated images — and proposes a principled-sounding solution via SVD-based subspace decomposition of VFM features. The empirical motivation is competent and the intuition behind freezing principal components to preserve semantic knowledge is at least coherent. However, the core theoretical claims are not substantiated at any rigorous level. The central assertion — that SVD's principal components correspond to 'semantic knowledge' while residual components correspond to 'forgery patterns' — is presented as…