Training-free AI-generated Image Detection via Spectral Artifacts

ICLR 2026 Conference Submission23896 Authors

Keywords: Training-free detection, Anomaly detection

Abstract: The rapid progress of generative models has enabled the synthesis of photorealistic images that are often indistinguishable from real photographs, raising serious concerns about misinformation and malicious use. While most existing AI-generated image (AIGI) detection methods rely on supervised training with labeled synthetic data, they struggle to generalize to unseen generators and incur substantial overhead for retraining. In this work, we propose SpAN, a simple yet effective training-free detection framework based on spectral analysis. Our key observation is that upsampling operations in generative models inevitably introduce spectral artifacts, which remain most pronounced at the axial Nyquist frequencies, even when images appear realistic. Building on this insight, we design two techniques to enhance detection reliability: (1) power calibration via azimuthal integration to mitigate bias from image-specific frequency distributions, and (2) autoencoder-based reconstruction to amplify residual artifacts and enable discrepancy-based scoring between original and reconstructed images. Extensive experiments across multiple datasets and generative models demonstrate that SpAN achieves robust and generalizable detection performance. For example, SpAN outperforms other training-free detection methods by a substantial margin (+0.241 AUROC) in the Synthbuster benchmark, which contains recent generative models.

Primary Area: generative models

Submission Number: 23896