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

Authors: Seunghyun Lee, Byoungkwon Kim, Jaehyun Nam, Kyungmin Lee, Jinwoo Shin

Published: 2026-06-07 13:56:17+00:00

Comment: Preprint. 22 pages, 10 figures, supplementary material included

AI Summary

This paper introduces SSAFE, a simple yet strong approach for AI-generated image detection using frozen multimodal vision encoders. It demonstrates that these encoders naturally separate real and synthetic images in their embedding space, enabling a linear classifier to achieve robust performance without extensive fine-tuning. The authors also propose a representation-aware data curation strategy that selects a compact, yet diverse, set of representative generators for training.

Abstract

The rapid advancement of generative models has blurred the boundary between synthetic and real imagery, creating an urgent need for reliable deepfake detection. Yet most existing approaches rely on massive real--fake datasets, which are increasingly difficult to maintain as new generators continue to emerge. In this work, we investigate how much information about image authenticity is already encoded in modern multimodal vision representations. We find that frozen multimodal encoders naturally separate real and synthetic images in their embedding space, enabling a simple linear classifier to achieve strong performance without task-specific fine-tuning. Motivated by this observation, we develop a representation-aware data curation strategy that selects a compact set of representative generators for training. The resulting training set contains only 10K images, compared to 288K in AIGIBench and 4M in OpenFake, while improving robustness to unseen generators and distribution shifts. We additionally introduce RealWorldBench, a benchmark consisting of modern camera photographs, contemporary stock images, and outputs from recent commercial generators. Experiments across multiple benchmarks show that combining frozen multimodal representations with carefully curated training data provides a simple and effective approach to AI-generated image detection.

Key findings

The study reveals that frozen multimodal vision encoders inherently encode sufficient information to separate real from AI-generated images, allowing a simple linear classifier to achieve state-of-the-art detection. Their representation-aware data curation strategy is highly effective, enabling superior generalization to unseen generators and real-world data with significantly smaller training datasets (e.g., 10K images compared to millions).

Approach

The method leverages frozen multimodal vision encoders, specifically PE-Core, to extract discriminative embeddings from images. A lightweight linear classifier is then trained on these frozen embeddings to classify images as real or AI-generated. A crucial component is a representation-aware data curation strategy that identifies and selects a compact, diverse set of representative generative models to build efficient training datasets.

Datasets

AIGIBench, AIGI-Holmes (including CNNDetection, GenImage, DRCT components), OpenFake, RealWorldBench (introduced by authors), Rapidata Human Preference Dataset, Pexels, iStockPhoto, Open Images V7, Pixabay, Unsplash (Kaggle High-Quality Face Dataset).

Model(s)

UNKNOWN

Author countries

South Korea

← Previous