SSAFE: Simple and Strong AI-Generated Image Detection via Frozen Vision Encoders

A new detection method called SSAFE uses frozen multimodal vision encoders and a compact 10,000-image training set to identify AI-generated images, according to research submitted in 2026. The approach avoids the massive datasets common in the field while improving robustness against unseen image generators. The rapid improvement of generative models has made it increasingly difficult to distinguish synthetic images from real photographs, creating demand for reliable detection tools [1]. Most current detection systems depend on large collections of real and fake images, which become harder to maintain as new image generators appear [1]. The researchers behind SSAFE examined how much authenticity information already exists inside modern multimodal vision representations. They found that frozen multimodal encoders naturally separate real and synthetic images in their embedding space, allowing a simple linear classifier to perform well without task-specific fine-tuning [1][2]. Based on that finding, the team built a representation-aware data curation strategy that selects a compact set of representative generators for training. The resulting training set contains only 10,000 images, compared to 288,000 in AIGIBench and 4 million in OpenFake [1][2]. Despite the smaller dataset, the method showed improved robustness to unseen generators and distribution shifts [1]. The researchers also introduced RealWorldBench, a new evaluation benchmark composed of modern camera photographs, contemporary stock images, and outputs from recent commercial generators [1][2]. Experiments across multiple benchmarks demonstrated that pairing frozen multimodal representations with carefully curated training data offers a straightforward path to AI-generated image detection [1][2]. The work arrives as synthetic media detection remains an active research area. Other recent papers on arXiv have explored related directions in computer vision and pattern recognition, though the SSAFE approach is notable for its reliance on frozen encoders rather than extensive fine-tuning [3][4][5]. The paper was submitted on June 7, 2026, to the Computer Science section on Computer Vision and Pattern Recognition [1].