DCCT: Demosaicing-guided Color Correlation Training

• The paper introduces a camera-aware, self-supervised DCCT framework that leverages demosaicing and CFA-induced color correlations to robustly separate photographs from AI-generated images.
• It demonstrates significant improvements over artifact-based detectors with cross-generator accuracies up to 97.25% and resilient performance under JPEG compression and downsampling.
• DCCT employs a U-Net architecture with a mixture-of-logistic parameterization and staged classifier training to enhance generalization across diverse generative pipelines.

Demosaicing-guided Color Correlation Training (DCCT) is a camera-aware, self-supervised learning framework designed to address the generalization failure of existing AI-generated image detectors that rely on generative artifacts. Leveraging the physical color correlations imposed by color filter arrays (CFA) and demosaicing in digital cameras, DCCT targets intrinsic distributional differences between photographs and synthetically generated images. This approach achieves robust and generalizable detection, substantially outperforming artifact-based detectors, particularly on unseen generators and under benign post-processing (Zhong et al., 30 Jan 2026).

1. Motivation and Context

Traditional approaches to detecting AI-generated images focus on sources of generative artifacts—such as upsampling artifacts in GANs or reconstruction errors in diffusion models. These methods are inherently specific to the generator architecture and deteriorate significantly when faced with novel or unseen generative pipelines. In contrast, all digital photographs undergo a physically constrained imaging pipeline involving CFA sampling, demosaicing, and subsequent imaging signal processing (ISP). The CFA, typically the Bayer RGGB pattern, samples only a single primary color per pixel; the remaining channels are reconstructed by demosaicing, imposing consistent inter-channel correlations and high-frequency aliasing signatures that current synthetic image generators, trained only on RGB statistics, fail to replicate. DCCT exploits this disparity by explicitly modeling color correlations rooted in camera hardware, establishing a feature space with a provable and generalizable separation between photographs and AI-generated images.

2. Color Filter Array Simulation and Conditional Modeling

An RGB image patch is denoted $x \in \mathbb{R}^{H \times W \times 3}$. The Bayer CFA sampling is simulated using binary masks $M_R, M_G, M_B \in \{0,1\}^{H \times W}$, marking spatial sampling positions for each channel within the Bayer pattern. The process forms a single-channel "mosaicked" observation:

$$x_c(i,j) = \sum_{c' \in \{R,G,B\}} M_{c'}(i,j) \cdot x_{c'}(i,j)$$

where $c$ indexes the observed CFA channel. The other two channels at each location act as reconstruction targets $t \in \mathbb{R}^{H \times W \times 2}$. DCCT is formulated as a self-supervised prediction task, modeling the conditional distribution

$p(t | x_c)$

where $t$ contains the missing channels and $x_c$ is the mosaicked input. In practice, the framework operates on high-pass filtered residuals $(x', y')$ to accentuate CFA-related aliasing subbands important for discrimination.

3. Architecture: Self-Supervised U-Net and Mixture of Logistic Parameterization

The core DCCT model is a U-Net architecture configured as an encoder-decoder with skip connections. The network receives high-pass filtered image stacks $x' \in \mathbb{R}^{H \times W \times M}$, where $M=30$ corresponds to rich-model (SRM) filters, optimized to accentuate inter-channel statistics induced by the CFA. The output parameterizes a mixture of $K=10$ logistic functions for each spatial location to predict $y'\in \mathbb{R}^{H\times W\times 2}$:

$$p(y'_{i,j}|x') = \sum_{k=1}^K \pi_{k,i,j}(x') \;\mathrm{Logistic}(y'_{i,j}; \mu_{k,i,j}(x'), s_{k,i,j}(x'))$$

where $\pi_{k,i,j}$ are the mixture weights, $\mu_{k,i,j}$ and $s_{k,i,j}$ are respectively the mean and scale of each logistic component. All parameters are regressed by the U-Net head. The network is trained on photographic samples $(x', y')$ under the negative log-likelihood:

$$\ell_\mathrm{NLL}(\theta) = -\sum_{i,j} \log p_\theta(y'_{i,j}|x')$$

For implementation, $H=W=64$, batch size is $16$, the optimizer is Adam with learning rate $1\,\text{e--4}$, and all inputs are truncated then rescaled to $[-1,1]$.

4. Theoretical Separation: Distributional Gap Analysis

A central result establishes that under reasonable Gaussian assumptions for $(x',y')$ and with sufficient high-pass energy, the 1-Wasserstein distance between photographic and AI-generated conditional distributions admits a uniform positive lower bound:

$$W_1\bigl(p(y'|x'),q(y'|x')\bigr) \geq \delta \quad \forall x' \text{ with aliasing energy}$$

The proof leverages Kantorovich–Rubinstein duality, showing the mean conditional prediction for photographs aligns with the CFA-induced transfer operator $T_\mathrm{CFA}$, while for synthesized images it aligns with a generator-dependent operator $T_\mathrm{Gen}$ that differs non-trivially at aliasing frequencies. This theoretical separation ensures the self-supervised DCCT task produces a feature space with an intrinsic, generator-agnostic gap between photographs and AI-generated images.

5. Classification Pipeline Construction

DCCT employs a two-stage supervised pipeline:

• Stage I: Two conditional models are trained independently: $f_\theta$ on photographs (minimizing $\ell_p$), and $f_\phi$ on AI-generated patches (minimizing $\ell_q$). Their weights are frozen post-training.
• Stage II: Feature maps (e.g., predicted means $\mu$ or full logistic parameterizations) from both $f_\theta$ and $f_\phi$ are concatenated and input to a shallow classifier $g_\psi$, consisting of a 4-block ResNet, a 2-layer Transformer, and a fully connected head. Cross-entropy loss over the binary photo/AI label is minimized. During inference, predictions are obtained by averaging over 16 random image crops.

This configuration exploits the full structural gap induced by the CFA-guided modeling in the U-Net backbone and consistently generalizes across unseen domains.

6. Experimental Results and Evaluation

DCCT demonstrates substantial improvements over artifact-based detectors:

• Cross-Generator Generalization: On the GenImage benchmark (8 generators), DCCT achieves 97.25% mean accuracy, outpacing the prior SOTA (Effort, 91.10%) and others (<90%), including robust performance (>92%) on mismatched test generators (BigGAN, ADM).
• Cross-Variant Generalization: Over DRCT-2M (16 diffusion variants), DCCT attains 91.13% overall accuracy versus DRCT's 90.49%, achieving >99% on most variants except on images with large untouched photo regions.
• Post-Processing Robustness: Under JPEG compression (QF $\in [70,100]$) and downsampling ($r \in [0.6,0.9]$), DCCT's accuracy decreases by less than 5%, significantly outperforming previous approaches.
• Forward Compatibility: On new generators (GigaGAN, DFGAN, GALIP, FLUX.1, SD-3.5-Turbo, Qwen-Image), DCCT achieves 98.79% accuracy, compared to 92.67% (Effort) and 48.66% (UnivFD).
• One-Class Anomaly Detection: The variant DCCT$^\dag$—trained only on photographs and deployed as an OOD detector using the statistic $D = \text{mean}(\text{NLL} - \text{entropy})$—achieves 88.36% accuracy on GenImage, outperforming most supervised baselines without exposure to synthetic images during training.

7. Limitations and Prospects

DCCT's current instantiation is tailored to the Bayer CFA and is challenged by mixed-content or diffusion reconstruction (DR) images where genuine photo regions dominate global color correlations. Extending the framework to alternative CFA designs (e.g., Fujifilm X-Trans), multi-spectral sensor arrangements, or smartphone-specific ISP variants represents a direct extension. Countermeasures by adversarial generators that explicitly reproduce CFA-like aliasing could reduce the efficacy of DCCT; addressing such cases may require integrating deeper ISP forensics or leveraging multi-modal cues, such as sensor noise properties or EXIF metadata.

By replicating the CFA simulation, high-pass filtering (30 SRM kernels), U-Net and mixture-of-logistic modeling, and staged classifier training as described, DCCT can be reproduced and adapted for diverse camera or generative contexts (Zhong et al., 30 Jan 2026).