CHROMA: Detecting AI-Generated Images through Inter-Channel Color-Space Correlations

Abstract: The rapid adoption of diffusion and large-scale generative models has made it increasingly challenging to distinguish synthetic imagery from real photographs. While automated detectors have been proposed, their generalization to unseen generators remains brittle. To address this limitation, we investigate inter-channel color correlations, a lightweight and underexploited forensic cue. We first demonstrate that LPIPS, a widely used perceptual metric, exhibits inconsistent responses to perturbations that selectively alter channel dependence across different color-space parameterizations, indicating that cross-channel statistics are not uniformly constrained by common perceptual training objectives. Motivated by this, we analyze the distributions of pairwise inter-channel correlation features across multiple color spaces. Our analysis reveals systematic, generator-specific differences in these distributions, with RGB and Lab color spaces providing the most apparent separation between real and generated images. Building on this, we introduce Chroma, a detector of AI-generated images which augments standard RGB inputs with inter-channel correlation maps and employs a fixed CNN backbone trained with a modest computational budget. We assess its robustness under both single-generator training and a limited multi-generator supervision regime, where only a few samples from additional generators are available. Across a standard benchmark protocol, correlation-augmented inputs improve real-vs-generated discrimination and robustness, yielding performance competitive with recent detectors while maintaining a simple architecture and training procedure. Code is available at https://github.com/JPSoteloSilva/CHROMA

• The paper presents a novel detector, CHROMA, that leverages inter-channel color-space correlations to differentiate real from AI-generated images.
• It employs a modified ResNet-50 with augmented correlation maps in RGB and Lab, achieving competitive AUC performance, particularly in Lab space.
• The study offers a practical, interpretable, and computationally efficient forensic approach that enhances robustness against unseen image generators.

Summary of "CHROMA: Detecting AI-Generated Images through Inter-Channel Color-Space Correlations"

Motivation and Context

The proliferation of diffusion-based and LLM-driven generative models has increased the difficulty of forensic analysis for distinguishing synthetic imagery from natural photographs. Existing detectors frequently overfit to generator-specific or dataset-specific characteristics, resulting in poor robustness to unseen generators. Prior approaches targeting low-level forensic cues—frequency-domain artifacts and marginal color statistics—have demonstrated partial transferability but often require complex preprocessing or large-scale multi-generator training. This paper posits that pairwise inter-channel correlation statistics in color spaces, especially RGB and Lab, encode generator-specific traces underexploited by both perceptual objectives and prior forensic methods. These traces are shown to exhibit systematic distributional shifts between real and synthetic images, providing a lightweight yet robust cue for detection.

Analytical Framework and Experimental Design

Color-Space Correlation Analysis

The authors begin by examining the sensitivity of LPIPS—a widely used perceptual similarity metric—to controlled perturbations that modulate inter-channel dependencies in various color spaces. The experiments reveal that LPIPS is not invariant to noise injected in different color-space parameterizations. Specifically, decorrelated noise in YUV induces significantly higher LPIPS distances than in RGB or Lab, and LPIPS responses are inconsistent regarding channel-correlation disruption, particularly in Lab versus RGB. These findings expose a lack of uniform constraint on cross-channel statistics by perceptual losses, motivating direct analysis of correlation descriptors.

Distributional Study of Correlation Features

By computing local Pearson correlation maps over $7 \times 7$ windows in RGB, HSV, Lab, and YUV spaces for real (RAISE-1k) and synthetic (Synthbuster) datasets, the paper demonstrates that generator-dependent distributional shifts manifest most prominently in RGB and Lab spaces. Quantitative Wasserstein distance analysis confirms that the separability between real and fake images, and among generators, is greatest for these two spaces. This result underscores the relevance of inter-channel correlations as generator-differentiable forensic features.

CHROMA Detector Architecture and Evaluation

Input Representation

CHROMA augments the standard RGB input with stacked inter-channel correlation maps, computed in RGB, Lab, or both. The detector is implemented by replacing the first convolutional layer in ResNet-50 to accommodate the increased channel dimensionality (3/6/9), while keeping the backbone and training procedure fixed. This design facilitates isolation of the contribution from correlation-augmented representations without introducing architectural confounds.

Training Protocol

Following the benchmark protocol in [7], the model is evaluated under both single-generator training (10k Latent Diffusion images) and a limited multi-generator regime, featuring modest additions from multiple generators (total 16,088 images). The test suite spans a broad family of GAN, diffusion, and commercial generators, with per-generator AUC used as the primary robustness metric.

Numerical Results and Claims

CHROMA achieves competitive detection performance, especially with Lab correlation-augmented inputs. Across most generator families, Lab correlations systematically improve AUC relative to RGB-only training, confirming their utility as a forensic cue. RGB correlation maps, in contrast, often fail to add discriminative power and can negatively affect performance when used alone or in conjunction with Lab, reinforcing the claim that color-space parameterization is crucial.

Augmented multi-generator training substantially improves robustness on unseen generators, and CHROMA remains competitive with or outperforms recent detectors on several GAN and diffusion families, despite using a much simpler architecture and lower training budget. It is explicitly noted that CHROMA's efficacy varies across generator families; some commercial tools (e.g., Adobe Firefly, DALL-E 3) are less separable via correlation cues, likely due to enhanced matching of natural color dependencies or post-processing pipelines. This is contrasted with detectors relying on spectral or embedding domain features.

Practical and Theoretical Implications

CHROMA demonstrates that inter-channel color-space correlations, particularly in Lab space, encode generator-dependent statistical traces largely unconstrained by current perceptual training objectives. Their explicit inclusion as input features yields robust detection competitive with more complex or computationally intensive methods and can be deployed efficiently with modest training infrastructure. This insight points to a practical pipeline for forensic detection that is interpretable, computationally lightweight, and well-suited for real-world variability.

Theoretically, the non-uniform constraint of cross-channel dependencies by perceptual objectives highlights an exploitable gap in generative modeling, suggesting avenues for both defense (robust detector design) and attack (improving synthetic realism by explicitly regularizing inter-channel statistics). CHROMA opens venues for learning non-linear forensic color spaces, integrating correlation cues with localization or multi-modal prediction systems, and scaling robustness via larger generator-diverse datasets.

Future Directions

The paper underscores several potential research directions:

• Fusion of correlation-augmented representations with spectral and embedding-based forensic cues for multi-view detection.
• Learning forensic color spaces optimized for discrimination instead of fixed parameterizations.
• Application of correlation descriptors for manipulation localization and calibration under domain and generator shift.
• Analyzing the impact of post-processing pipelines in commercial tools on color correlation statistics.

Conclusion

"CHROMA: Detecting AI-Generated Images through Inter-Channel Color-Space Correlations" establishes the efficacy of inter-channel color correlations, particularly those derived from the Lab color space, as a robust and interpretable forensic cue for AI-generated image detection. The method attains competitive robustness with modest training budgets and architectures and highlights the importance of representation parameterization when leveraging color statistics. Future integration of correlation cues with complementary forensic signatures and the pursuit of learned color-space representations are promising directions for advancing synthetic image forensics and detection reliability (2606.08864).