Chromatic Prior-Guided Color Compensation

• Chromatic prior-guided color compensation methods explicitly leverage hand-crafted statistical or physics-inspired priors to model expected color distributions and correct distortions.
• These techniques integrate priors as pixel-wise transforms, adaptive fusion modules, or neural conditioning inputs, enabling effective restoration in underwater imaging, artwork recovery, and in-camera color processing.
• Empirical evaluations demonstrate improved UCIQE, UIQM, and ΔE2000 metrics, confirming that the hybrid approach of using priors with deep networks enhances both physical and perceptual color fidelity.

Chromatic Prior-Guided Color Compensation is an umbrella term for methodologies that explicitly leverage hand-crafted, statistically- or physics-inspired chromatic priors to guide the correction of color distortions in degraded images. These priors encode expected hue, chromaticity, or color ratios, either globally or locally, and are integrated into compensation pipelines—either as preprocessing modules, conditioning inputs for neural networks, or via adaptive blending rules. The approach is particularly impactful in scenarios where data-driven models alone struggle to restore physically plausible chromatic cues, such as underwater imaging, color-degraded artwork restoration, or scenes under non-ideal illumination. Chromatic prior-guided color compensation has seen varied implementations, ranging from classical pixel-wise transforms to modules within advanced diffusion and transformer architectures.

1. Mathematical Formulation of Chromatic Priors

Chromatic priors characterize assumptions about physically plausible or statistically likely color distributions for given imaging conditions and are typically defined in a perceptually uniform color space (such as CIELAB) or chromaticity spaces.

In underwater imaging, a canonical example is the three-channel compensation (3C) method as utilized in AquaDiff (Shaahid et al., 15 Dec 2025). Given degraded RGB input $I$, the image is converted to Lab space to separate luminosity $L$ and chrominance $(a^*,b^*)$. Compensated chrominance channels are then computed: $I_a^c(x) = I_a(x) - \kappa\,M(x)\,G[I_a(x)]$

$I_b^c(x) = I_b(x) - \lambda\,M(x)\,G[I_b(x)]$

where $G[\cdot]$ denotes a small-kernel Gaussian blur, $M(x)$ is a binary-then-soft spatial mask suppressing adjustment in high-luminance regions, and $\kappa,\lambda$ are empirically chosen scalars (typically $\kappa=\lambda\approx 0.7$). The final color-compensated image $\mathbf{y}$ is obtained by recombining the original $L$ and compensated chrominance and inverse-transforming to RGB: $$\mathbf{y} = \mathcal{T}_{\mathrm{Lab}^{-1}}\bigl(L(x),\, \mathbf{c}(x)-(\boldsymbol\Lambda \odot M(x) \odot G[\mathbf{c}(x)])\bigr)$$ where $\mathbf{c}(x) = [a^*(x)\;b^*(x)]^\top$ and $\boldsymbol\Lambda = [\kappa,\,\lambda]^\top$ (Shaahid et al., 15 Dec 2025).

In adaptive underwater frameworks (Tian et al., 5 Mar 2025), multiple chromatic priors are instantiated:

• Red Channel Prior (RCP): boosts suppressed reds by adding a scaled $(I_G - I_B)$ term to the red channel.
• Dark Channel Prior (DCP) and Multi-Scale Dark Channel Prior (MUDCP): use minimum intensity statistics to estimate transmission and recover color.

For in-camera color correction under non-Planckian illumination (Tedla et al., 21 Nov 2025), prior information is parameterized by 2D chromaticity (CIE-xy or sensor-derived ratios) rather than traditional 1D correlated color temperature, improving compensation accuracy for contemporary light sources.

2. Color Compensation Transforms and Fusion Mechanisms

Chromatic prior-guided color compensation methods implement the prior either as an explicit pixel-wise transform, an adaptive fusion of multiple priors, or as a conditioning input for neural restoration. Three chief paradigms emerge:

Pixel-wise Compensation: In the 3C method, a simple but data-informed transform is applied in Lab space per pixel (Shaahid et al., 15 Dec 2025). In the ACC module (Tian et al., 5 Mar 2025), the output is a convex combination of RCP, DCP, and MUDCP estimates, with fusion weights $\boldsymbol{\alpha}$ determined by a function of per-channel attenuation and water-type index: $$J_{\mathrm{ACC}}(x) = \sum_{i=1}^3 \alpha_i\,J_i(x), \quad [\alpha_1, \alpha_2, \alpha_3] = \mathrm{softmax}(f(\eta; \mathrm{WTI}))$$ where $J_i$ are the outputs of individual priors, $\eta$ are normalized channel means, and $\mathrm{WTI}$ identifies the water type (Tian et al., 5 Mar 2025).

Demultiplexed Priors: In artwork restoration, localized color priors are extracted from masks on the faded ab channels, providing spatially precise compensation cues which guide transformer attention (Tang et al., 3 Nov 2025).

Chromaticity-Guided CST Prediction: For in-camera color processing, per-image or per-pixel 2D chromaticity features are input to an MLP that predicts the color space transform (CST) matrix to correct the raw RGB to display or XYZ space (Tedla et al., 21 Nov 2025).

3. Integration into Deep Generative and Discriminative Networks

When correcting color in severely distorted images, direct learning is often insufficient; thus, priors are incorporated as auxiliary conditioning signals or fusion sources within complex neural networks.

In AquaDiff (Shaahid et al., 15 Dec 2025), the color-compensated image $\mathbf{y}$ (from the 3C transform) serves as a persistent conditioning image, accessed by the reverse diffusion network via cross-attention at each denoising step. The cross-attention module computes, at each feature scale: $$\mathrm{CrossAtt}(x_t, y) = \mathrm{Softmax}\left( \frac{Q(x_t) K(y)^\top}{\sqrt{d_k}} \right) V(y)$$ where $Q(x_t)$ are query features, $K(y)$ and $V(y)$ are key/value projections from the color-compensated image. This dynamic fusion allows the backbone (an enhanced U-Net with residual dense and multi-resolution attention blocks) to selectively incorporate prior-guided chromatic structure throughout the generation process (Shaahid et al., 15 Dec 2025).

Analogously, in PRevivor (Tang et al., 3 Nov 2025), hue correction is mediated by a dual-branch cross-attention module: one branch is strictly masked by localized hue priors, while a second global branch reasons unconstrained, allowing both precise local enforcement and global adaptation in restoration of ancient artwork. The masked attention uses a hard block (large negative logits) where the prior mask $M_{\text{prior}}$ is zero, ensuring that guidance is only active for confident regions.

In in-camera color pipelines (Tedla et al., 21 Nov 2025), chromatic priors derived from 2D chromaticity features parametrize a lightweight MLP that predicts the CST for color mapping shot-by-shot or pixel-wise, supporting fast adaptation to varying and mixed illuminant fields.

4. Training Strategies and Loss Functions

Supervision strategies depend on the modality and objective. In most deep generative paradigms, the prior is not supervised directly. For example:

• In AquaDiff (Shaahid et al., 15 Dec 2025), there is no explicit loss on the compensated chromatic channels. The only loss is a joint cross-domain consistency loss over the final output, comprising pixel-level $\ell_1$, multi-scale $\ell_1$, perceptual (VGG), SSIM, and frequency-domain magnitude terms: $$\mathcal{L}_{\mathrm{CDC}} = \frac{1}{HWC}\|\hat x_0 - x_0\|_1 + \ldots + \sum_l \frac{w_l}{H_lW_lC_l}\|\phi_l(\hat x_0)-\phi_l(x_0)\|_2^2 + \ldots$$ This ensures the network preserves both structural and chromatic cues delivered via the prior-conditioned guidance image.
• In prior-guided restoration of artworks (Tang et al., 3 Nov 2025), the loss for hue correction combines pixel, adversarial, perceptual, and a masked pixel loss that correlates the corrected ab channels to the localized hue prior: $$L_{\text{hue}} = \lambda_{pix}\|y_{ab} - \hat y_{ab}\|_1 + \lambda_{mask}\|M_{prior} \odot (y_{ab}-\hat y_{ab})\|_1 + \ldots$$ enforcing color fidelity especially in regions with confident prior information.
• In in-camera color mapping (Tedla et al., 21 Nov 2025), cosine loss on color vectors (stable proxy for angular error) is used for optimizing the CST-MLP, with additional robustness provided by noise augmentation on the chromaticity features.

Hybrid classical pipelines (e.g., (Tian et al., 5 Mar 2025)) do not learn parameters for the chromatic priors themselves, instead relying on algorithmic outputs and evaluating final performance with UCIQE, UIQM, CIEDE2000, and SSIM.

5. Empirical Impact and Quantitative Performance

The empirical advantages of chromatic prior-guided compensation are consistently observed:

• In underwater scenarios, the 3C prior in AquaDiff yields higher UCIQE color fidelity scores and more robust correction under extreme attenuation compared to pure concatenation-conditioning, standard CNNs, or GAN-based methods (Shaahid et al., 15 Dec 2025).
• Adaptive blending of classic chromatic priors in ACC (RCP, DCP, MUDCP) (Tian et al., 5 Mar 2025) results in elevated UIQM and UCIQE, with perceptual color differences (CIEDE2000) systematically improved across diverse water types.
• For in-camera color correction, using 2D chromaticity as a prior (rather than 1D CCT) and leveraging an MLP-based CST mapping achieves a 22% reduction in average angular error under challenging off-locus LED illumination and substantial improvements in $\Delta E_{2000}$ (from $\sim$7.21 to $\sim$6.60) without sacrificing cost or performance on classical light sources (Tedla et al., 21 Nov 2025).
• In artwork restoration, explicit imposition of localized hue priors leads to superior $\Delta$Colorfulness reduction (1.70 vs 3.92 compared to BigColor) and competitive SSIM and FID on test sets (Tang et al., 3 Nov 2025).

These results demonstrate that feeding strong, interpretable chromatic cues to color enhancement models systematically drives higher physical and perceptual color fidelity in domains where generative models alone cannot reliably infer plausible color distributions.

6. Applications and Limitations

Chromatic prior-guided color compensation has a wide applicability in:

• Underwater imaging, where wavelength-dependent attenuation and scattering cause severe red/green suppression and blue-green dominance.
• Digital restoration of color-degraded paintings and artifacts, where lost chromatic cues are only partially preserved in the faded structure and must be inferred using prior distributions or historical exemplars.
• In-camera color pipelines for modern digital photography, addressing the deficiencies of classical color temperature-based methods in the face of modern, off-Planckian LED illumination.

Nevertheless, these methods depend critically on the informativeness and suitability of the chosen prior for the given domain. For scenes outside the support of the prior (e.g., exotic color distributions, unmodeled illumination spectra), compensation may be suboptimal. A plausible implication is that future systems will increasingly incorporate meta-learning or domain-adaptive adjustment of priors for further robustness. Another consideration is that, while priors accelerate convergence and improve plausibility, over-reliance on rigid priors can suppress rare or contextually valid colorations.

7. Connections to Related Paradigms and Future Directions

Chromatic prior-guided compensation interacts with, but is distinct from, fully data-driven end-to-end learning. Methods such as color constancy via chromaticity-luminance histograms (Chakrabarti, 2015) similarly leverage global or empirical priors for illuminant estimation but do not typically provide pixel-wise guidance for generative restoration. The trend across recent work is toward hybrid models that integrate the interpretability and domain knowledge of priors with the adaptivity and capacity of modern neural networks (Shaahid et al., 15 Dec 2025, Tang et al., 3 Nov 2025).

Future directions may include automated or learned prior extraction, context-aware or physics-based prior selection, and unified frameworks that allow priors to be updated online as new, domain-specific information accrues. The continued theoretical integration of physical models, statistics, and learning-based restoration remains an active and fruitful research area.