The paper "RIGID: A Training-Free and Model-Agnostic Framework for Robust AI-Generated Image Detection" introduces a novel method called RIGID (Robust AI-Generated Image Detection) to distinguish between real and AI-generated images. The authors observe that real images are more robust to small noise perturbations in the representation space of vision foundation models, which is a key insight for their approach. RIGID is designed to be training-free and model-agnostic, meaning it does not require large datasets or specific knowledge of the generation process. The method compares the cosine similarity between the original and noise-perturbed images to detect AI-generated content. Evaluations on diverse datasets and benchmarks show that RIGID outperforms existing training-based and training-free detectors, with an average performance improvement of over 25% compared to the best training-free method, AEROBLADE. RIGID also demonstrates strong generalization across different image generation methods and robustness to image corruptions. The paper discusses the limitations of training-based and training-free methods and highlights the practical and robust nature of RIGID, addressing the growing concerns about the misuse of generative AI technology.The paper "RIGID: A Training-Free and Model-Agnostic Framework for Robust AI-Generated Image Detection" introduces a novel method called RIGID (Robust AI-Generated Image Detection) to distinguish between real and AI-generated images. The authors observe that real images are more robust to small noise perturbations in the representation space of vision foundation models, which is a key insight for their approach. RIGID is designed to be training-free and model-agnostic, meaning it does not require large datasets or specific knowledge of the generation process. The method compares the cosine similarity between the original and noise-perturbed images to detect AI-generated content. Evaluations on diverse datasets and benchmarks show that RIGID outperforms existing training-based and training-free detectors, with an average performance improvement of over 25% compared to the best training-free method, AEROBLADE. RIGID also demonstrates strong generalization across different image generation methods and robustness to image corruptions. The paper discusses the limitations of training-based and training-free methods and highlights the practical and robust nature of RIGID, addressing the growing concerns about the misuse of generative AI technology.