Synthetic Image Detection via Curvature of Diffusion Probability Flows

Jiaming Li, Weihua Ou, Ni Li, Meilin Zheng

Keywords: Synthetic Image Detection

Abstract: Synthetic image detection (SID) faces two major challenges: high computational cost from reconstruction-based methods and insufficient generalization. To address these issues, we propose a SID paradigm that leverages the ODE formulation of diffusion models. Instead of reconstructing images, our method analyzes probability flow trajectories from data distributions toward a Gaussian prior. We theoretically relate discrete step distances on the Wasserstein manifold to the kinetic energy of the probability flow. We further show empirically that trajectory deviation statistics derived from these distances correlate with reconstruction error and that real and synthetic images differ most in the early half of the diffusion inversion. In this regime, real images tend to exhibit higher curvature variance with occasional extreme deviations, whereas synthetic ones follow smoother and more consistent trajectories. Building on this observation, we introduce curvature features of probability flow trajectories as a discriminative signal for SID. To the best of our knowledge, this is the first work to exploit probability flow curvature for this task. Extensive experiments demonstrate that our method generalizes robustly to unseen models and achieves state of the art results across multiple benchmarks, while requiring less than half the FLOPs of reconstruction based detectors that perform full diffusion inversion.

Primary Area: generative models

Submission Number: 6224