Beyond Familiar Domains: A Study of the Generalization Ability of Machine-Generated Image Detectors

Modern generative models produce images that are nearly indistinguishable from human-created ones, posing serious challenges for content verification. As machine-generated content is increasingly integrated into professional workflows, the task of reliably detecting it becomes critically important. Existing detectors of machine-generated images do not generalize well to new generative models and visual domains. This paper investigates the ability of current detectors of machine-generated images to recognize new generative models and images from various domains not represented in the training data. The objects of the study include popular architectures, such as a combination of a pretrained CLIP with an MLP classifier and a model based on a mixture of experts. Particular attention is paid to analyzing the current limitations and reliability of both closed and open solutions, especially in the context of emerging new generative methods and specific types of images. Experimental results demonstrate significant limitations of existing approaches: models exhibit low generalization ability not only to new generators, but also when working with images from new domains.