Reproduce “LATTE: Latent Trajectory Embedding for Diffusion-Generated Image Detection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ana Vasilcoiu, Ivona Najdenkoska, Zeno Geradts, Marcel Worring
LATTE uses selected latent states from diffusion denoising, fusing each with image features for classification. Unlike LaRE’s single-step reconstruction-error feature, it combines a sequence of model responses. On GenImage, trained on SDv1.4 images, it reaches 91.0% average accuracy across eight generator subsets, versus 86.9% for AIDE; this benchmark does not establish performance on arbitrary future generators.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.