CoDA: Color Distribution Probing for Efficient and Generalizable AI-Generated Image Detection

Technical Summary: CoDA: Color Distribution Probing for Efficient and Generalizable AI-Generated Image Detection

Problem Statement
AI-generated image detection currently faces a fundamental trade-off between generalization and efficiency. Lightweight, artifact-based methods often degrade when encountering unseen generators or visual domains, while robust large-scale models (e.g., Vision-Language Models) are computationally expensive and may still struggle with underrepresented domains like medical or technical imagery. Furthermore, existing benchmarks primarily focus on cross-model evaluation within the narrow regime of photorealistic imagery, leaving cross-domain robustness underexplored. This paper addresses the need for a detection paradigm that is both computationally efficient and robust across diverse generators and visual domains.

Methodology
The authors propose CoDA, a compact detection framework (1.48M parameters) built on the observation that AI-generated images often exhibit characteristic color imbalances compared to real photographs. The methodology consists of three core components:

1. Theoretical Basis (Manifold Divergence): The paper posits that generative models optimize proxy objectives that can push samples off the low-dimensional manifold of real images. This deviation is particularly sensitive in chromatic directions due to the lack of physical constraints (e.g., camera ISP pipelines like white balance and gamma correction) in generative processes. Theoretical analysis links this off-manifold deviation to color distribution non-uniformity.
2. Noise-Quantization Probe: To quantify color distribution stability, the authors introduce a probe that applies controlled stochastic perturbation. The process involves:
   • Injecting Gaussian noise into the input image.
   • Applying color quantization (rounding).
   • Averaging multiple restored samples.
   • Computing the difference map between the original and the restored image.
   • Mechanism: Real photographs, with smoother and broader color distributions, tend to yield perturbation errors that cancel out during averaging, resulting in small residuals. Conversely, synthetic images with peaked or irregular color distributions produce systematic, structured residuals.
3. Dual-Branch Architecture: CoDA employs a lightweight dual-branch network.
   • Visual Branch: Processes the original image to capture semantic and textural features.
   • Color Branch: Processes the difference map generated by the Noise-Quantization Probe to extract distributional cues.
   • The features from both branches are concatenated and fed into a lightweight classifier. This design ensures the model does not rely solely on color but uses it as a complementary signal to enhance robustness.

Key Contributions

• FakeForm Benchmark: The authors introduce FakeForm, a large-scale benchmark comprising approximately 370,000 images across 62 diverse domains (including CT scans, depth maps, flowcharts, and artistic media). It supports both cross-model and cross-domain evaluation and includes over 760,000 human perceptual judgments with realism scores and textual justifications.
• Color-Distribution Analysis: The paper identifies and analyzes color-distribution artifacts as a recurring signal across major generation paradigms (GANs, Diffusion, Flow-based, and Autoregressive models), providing a theoretical link between manifold divergence and color non-uniformity.
• CoDA Detector: A lightweight detector that transforms color-distribution irregularities into robust signals via the Noise-Quantization Probe, achieving high efficiency without sacrificing generalization.

Experimental Results

• Standard Benchmarks: On established benchmarks (ForenSynths, Ojha, GenImage), CoDA achieves state-of-the-art performance, outperforming or matching leading methods like FatFormer and SAFE. For instance, on the GenImage benchmark, CoDA achieves 95.9% accuracy and 99.1% AP.
• Cross-Domain Generalization (FakeForm): CoDA demonstrates superior robustness in cross-domain settings. While state-of-the-art detectors often drop below 75% accuracy on challenging cross-domain content, CoDA achieves 77.7% average accuracy across 62 domains. It significantly outperforms large VLM-based baselines (e.g., AIGI-Holmes) which degrade more substantially under domain shifts.
• Efficiency: CoDA processes images at 125.2 frames per second (FPS) with only 1.48M parameters. In contrast, VLM-based systems like AIGI-Holmes operate at roughly 2.2 FPS with over 7B parameters.
• Robustness: The model maintains high performance under common image perturbations (Gaussian blur, cropping, JPEG compression, noise), showing smaller performance drops compared to competitors, particularly on diffusion-generated images.
• Ablation Studies: Experiments confirm the complementarity of the two branches. The full model outperforms the "Image Branch Only" (66.8% vs. 77.7% on FakeForm) and "Color Branch Only" configurations, validating that color cues provide essential complementary evidence, especially in color-rich domains.

Significance and Claims
The paper claims that persistent generative artifacts, specifically color distribution non-uniformity, provide a practical foundation for efficient and robust AI-generated image detection. By leveraging the Noise-Quantization Probe, CoDA occupies a "middle ground" between brittle low-level forensic methods and expensive high-level semantic models. The authors argue that their approach offers a scalable solution for real-world deployment where detectors must operate across a broad spectrum of visual content (from photography to technical diagrams) without the computational overhead of large multimodal models. The work suggests that while color cues are domain-dependent (performing best in photographic and artistic content and less so in low-color or highly specialized domains), they serve as a powerful, complementary signal that enhances generalization when fused with standard image features.