Can Vision Mamba beat CNNs and ViTs at detecting fake images?

As AI-generated imagery becomes increasingly realistic, detecting fake visuals is a critical challenge for fighting misinformation and protecting privacy. Researchers Mamadou Keita, Wassim Hamidouche, and colleagues from multiple institutions have conducted an in-depth investigation into whether Vision Mamba—a relatively new architecture inspired by state-space models—can improve detection of AI-generated images. The study, posted on arXiv (arXiv:2605.14799), systematically compares multiple Vision Mamba variants against established methods including Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), and Vision-Language Model (VLM) based detectors.

The team evaluated performance across diverse datasets and synthetic image sources (from GANs, diffusion models, etc.), measuring accuracy, efficiency, and generalizability—the ability to detect images from unseen generators. Results show that Vision Mamba offers competitive accuracy in many scenarios, especially where efficiency matters, but still lags behind top ViT and VLM-based detectors in some benchmarks. The authors conclude that Vision Mamba shows promise as a lightweight component in hybrid detection systems, but is not yet a standalone replacement for current state-of-the-art methods. This research provides a crucial benchmark for future work in AI-generated image forensics.