Synthetic Image Detection via Curvature of Diffusion Probability Flows

ICLR 2026 Conference Submission6224 Authors

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 novel SID paradigm that leverages the ODE formulation of diffusion models. Rather than reconstructing images, our method analyzes the probability flow trajectories from data distributions to a Gaussian prior. We show that the discrete-step distances on the Wasserstein manifold inherently encode reconstruction error, and that real and synthetic images diverge most significantly in the early half of the diffusion inversion. Real images exhibit higher curvature variance with extreme deviations, whereas synthetic ones follow smoother, more consistent trajectories. Building on this insight, we introduce curvature features of probability flow trajectories as a new discriminative signal. To the best of our knowledge, this is the first work to exploit probability flow curvature for SID. Extensive experiments demonstrate that our method generalizes robustly to unseen models, achieves SOTA results across multiple benchmarks, and does so with less than half the computational cost of full diffusion inversion.

Primary Area: generative models

Submission Number: 6224