CINEMAE: Leveraging Frozen Masked Autoencoders for Cross-Generator AI Image Detection

Minsuk Jang, Hyeonseo Jeong, Minseok Son, Changick KimarXiv 2025 · DOI · arXiv

Introduction

As generative AI models like GANs and Diffusion models continue to evolve, the microscopic "artifacts"—spectral patterns or pixel-level inconsistencies—that traditional detectors rely on are rapidly vanishing. This creates an arms race where detection strategies must shift from hunting for residual traces to evaluating contextual plausibility. The core question becomes: "Does this specific part of the image make sense given its surroundings?"

Methodology: Contextual Image Anomaly Estimation

The proposed framework, CINEMAE (Contextual Image Anomaly Estimation with Masked AutoEncoder), leverages a pre-trained, frozen Masked AutoEncoder (MAE) as a "natural image prior." Because an MAE is trained to reconstruct missing patches based on visible ones, it inherently learns the statistical and semantic structures of natural images.

Overview of the CINEMAE architecture compared to text-based DetectGPT, highlighting the masking and reconstruction process.

The architecture employs two parallel branches:

  1. Semantic Branch: The MAE encoder extracts a global feature vector (the [CLS] token) to capture high-level semantic information.
  2. Anomaly Branch: This branch masks random patches and uses the MAE decoder to reconstruct them. It calculates a Local Contextual Anomaly Score (LCAS) by combining statistical deviation and reconstruction fidelity ($L_2$ distance).

Crucially, the authors found that keeping the MAE backbone frozen is superior to fine-tuning. A frozen backbone preserves the general "natural image prior" learned during ImageNet pre-training, preventing the model from overfitting to the specific artifacts of a single generator.

Experimental Results

CINEMAE was evaluated on massive benchmarks including GenImage (1 million images), AIGCDetectBenchmark (16 models), and the challenging Chameleon dataset.

The Predictability Gap

A key insight of this work is the "predictability gap." AI-generated images are mathematically more "regular" and easier for an MAE to reconstruct than real-world images, which contain higher entropy and more complex local statistics. Consequently, AI images consistently show lower reconstruction error and anomaly scores (approximated Negative Log-Likelihood).

The predictability gap across different generators (BigGAN, Wukong, Midjourney, SD v1.4) showing real images consistently have higher anomaly scores than AI-generated ones.

Generalization and Robustness

CINEMAE achieved state-of-the-art performance with a mean accuracy of 96.63% on GenImage. Notably, when trained on Stable Diffusion v1.4, it maintained over 95.9% accuracy across eight different generators, demonstrating exceptional cross-generator generalization.

Furthermore, the model is remarkably resilient to JPEG compression. While compression often destroys the high-frequency artifacts used by other detectors, CINEMAE’s reliance on semantic context allows it to maintain over 93% accuracy even at a heavy JPEG Quality Factor of 50.

Conclusion and Future Work

While CINEMAE sets a new standard for robust AI image detection, it faces challenges with low-context images (e.g., plain walls) where natural simplicity might be mistaken for AI regularity. Future work will explore more advanced context-modeling strategies and test the system's resilience against adaptive adversarial attacks. By focusing on contextual coherence rather than fleeting artifacts, CINEMAE provides a more sustainable path forward for identifying synthetic content.

Summary updated

BibTeX
@misc{jang2025cinemae,
  author = {Minsuk Jang and Hyeonseo Jeong and Minseok Son and Changick Kim},
  title = {{CINEMAE: Leveraging Frozen Masked Autoencoders for Cross-Generator AI Image Detection}},
  year = {2025},
  eprint = {2511.06325},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.06325},
}
Figure