NeXT-IMDL: A Benchmark for Robust Image Manipulation Detection

This paper addresses the critical need for robust Image Manipulation Detection and Localization (IMDL) methods in the face of increasingly accessible AI-generated content. It highlights the limitations of current evaluation methods, which often overestimate model performance due to their simplified cross-dataset approach. The paper's significance lies in its introduction of NeXT-IMDL, a diagnostic benchmark designed to systematically probe the generalization capabilities of IMDL models across various dimensions of AI-generated manipulations. This is crucial because it moves beyond superficial evaluations and provides a more realistic assessment of model robustness in real-world scenarios.