With the advancement of diffusion-based image generators, distinguishing between generated and real images is becoming increasingly difficult. This paper proposes a novel approach that leverages diffusion denoising cues but considers the sequential nature of the denoising process rather than relying on a single-step reconstruction. LATTE (LATent Trajectory Embedding) models the evolution of latent embeddings across multiple denoising stages, revealing subtle and discriminatory patterns that distinguish generated and real images. It demonstrates outstanding performance on benchmarks such as GenImage, Chameleon, and Diffusion Forensics, particularly in cross-generator and cross-dataset scenarios.
