Training-Free Detection of AI-Generated Images Based on Coherence Analysis of Image Representations
TL;DR
This work proposes a training-free framework for AI-generated image detection based on perturbation-induced representation coherence, and achieves the highest mean AUROC among the evaluated training-free methods, while remaining competitive with training-based approaches.
Abstract
Recent advances in generative AI (artificial intelligence) have enabled the creation of highly realistic visual content, posing significant challenges to image authenticity verification. Existing detection methods predominantly rely on supervised learning of generator-specific artifacts, potentially limiting their ability to generalize to unseen generation models and image manipulation scenarios. In this work, we propose a training-free framework for AI-generated image detection based on perturbation-induced representation coherence. Instead of measuring the response to a single perturbation, we construct multiple complementary residual-guided perturbation families and analyze the internal coherence of their feature representations extracted by a frozen DINOv3 model. Furthermore, we extend the proposed coherence analysis across network depth by introducing a layer-wise coherence trajectory, capturing the evolution of representation consistency throughout the hierarchical feature space. Extensive experiments on both fully AI-generated images and images locally modified using AI models demonstrate that the proposed framework achieves the highest mean AUROC among the evaluated training-free methods, while remaining competitive with training-based approaches.