The rapid advancement of diffusion-based image generators has made it increasingly difficult to distinguish generated from real images. This can erode trust in digital media, making it critical to develop generalizable detectors for generated images. Recent methods leverage diffusion denoising cues, but mainly focus on single-step reconstruction errors, ignoring the inherent sequential nature of the denoising process. In this work, we propose LATTE - Latent Trajectory Embedding - a novel approach that models the evolution of latent embeddings across several denoising timesteps. By modeling the trajectory of such embeddings rather than single-step errors, LATTE captures subtle, discriminative patterns that distinguish real from generated images. Each latent is refined by employing our latent-visual feature refinement module and aggregated into a unified representation. Afterwards, it is fused with the visual features and finally passed into a lightweight classifier. Our experiments demonstrate that LATTE surpasses the baselines on several established benchmarks, such as GenImage and DiffusionFake. Moreover, it demonstrates strong performance in cross-generator and cross-datasets settings, highlighting the potential of using the trajectory of latent embeddings for generated image detection. The code is available on the following link: this https URL.
View on arXiv