FusionDetect: New Framework for Fake Image Detection Across Generators and Domains

Researchers have introduced FusionDetect, a novel method for detecting synthetic images, alongside the OmniGen Benchmark, a comprehensive evaluation dataset. The work, detailed in arXiv paper 2510.05740, argues that current detection efforts focus too narrowly on cross-generator generalization, overlooking the equally important challenge of generalization across visual domains. FusionDetect leverages two frozen foundation models, CLIP and Dinov2, to create a cohesive feature space that adapts to changes in both content and design. The OmniGen Benchmark incorporates 12 state-of-the-art generators to provide a realistic evaluation of detector performance. This research addresses the rapid development of generative models, emphasizing the need for reliable detectors.