Title: Chroma Clues: Leveraging Color Statistics to Detect Synthetic Images

Abstract:

The proliferation and spread of AI-generated imagery are accelerating at a historic pace. As image synthesis models rapidly advance toward the ability to flawlessly replicate natural scenes, they pose significant challenges to the field of image forensics. This study investigates a previously overlooked vulnerability in contemporary generative models: their difficulty in accurately reproducing the color statistics characteristic of real-world images.

We begin by demonstrating that the LPIPS loss function, commonly employed in training image generators, exhibits lower sensitivity to chrominance compared to luminance. This imbalance likely results in statistical anomalies within the color distribution of synthetic outputs. Leveraging this finding, we propose six hand-crafted color transformations alongside a technique for learning a task-specific color transform designed to statistically highlight generated content.

These transformations offer multiple applications. First, we extract color-sensitive features at both the pixel and patch levels. Utilizing these features, a straightforward and interpretable classifier achieves an average generalization accuracy of 93.27%, while maintaining strong resilience against six distinct types of post-processing. Second, we show that these transformations reveal distinct visual noise patterns that differentiate natural from synthetic regions, facilitating intuitive visual assessment. Finally, we illustrate how these transforms can accentuate color patterns in generated images, thereby improving multiclass attribution capabilities.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC