Chromatic Prior-Guided Conditioning

• Chromatic Prior-Guided Conditioning is a set of techniques that inject explicit color distribution constraints into deep models to enhance color fidelity and semantic alignment.
• The approach leverages methods like cross-attention, latent-space fusion, and classifier-free guidance to adaptively integrate chromatic priors during model inference.
• Empirical results show that CPGC significantly improves performance in tasks such as color constancy, text-to-image synthesis, underwater enhancement, and automatic colorization.

Chromatic prior-guided conditioning (CPGC) refers to a family of techniques in image synthesis, enhancement, and color constancy in which explicit statistical or structural constraints on color distributions—referred to as chromatic priors—are injected into models as conditioning signals. These priors inform the learning and inference process of deep generative or discriminative methods, improving color plausibility, semantic alignment, or restoration accuracy. Modern CPGC architectures leverage cross-attention mechanisms, latent-space fusion, or classifier-driven aggregation to ensure that chromatic characteristics such as palette, illumination, or structural color cues are respected at every stage of the computation, leading to substantial improvements over naive or unconditioned approaches.

1. Chromatic Priors: Definition and Characterization

Chromatic priors are data-derived or analytically designed summaries of color distributions expected in a given imagespace, domain, or context. Formally, these priors can take the form of histograms (e.g., over CIE L*a*b* color space), pixelwise luminance-chromaticity joint distributions, compensated chromatic channels for degraded images, or statistical models of illuminants. The form and extraction protocol for a chromatic prior are application-dependent:

• In color constancy (Chakrabarti, 2015), the per-pixel prior is an empirical or learned log-likelihood $L[\hat{x}, y]$ for chromaticity $\hat{x}$ conditioned on luminance $y$.
• In text-to-image generative modeling (Aggarwal et al., 2023), the prior is a color histogram $z_c$ over L*a*b*, projected to match the latent dimension of CLIP embeddings and used as a token in a diffusion prior.
• In underwater image enhancement (Shaahid et al., 15 Dec 2025), the prior is a compensated two-channel a/b map $I^{c}_a(x), I^{c}_b(x)$ correcting for wavelength-dependent loss before conditioning the denoiser.
• In automatic colorization (Wang et al., 2024), luminance and higher-level semantic maps act as "chromatic priors" entering the diffusion process and final decoder.

These representations encode either global (image-wide) statistics, local (spatially varying) color structure, or both.

2. Methodologies for Chromatic Prior-Guided Conditioning

Several methodological paradigms exist for utilizing chromatic priors to condition deep models:

A. Empirical and Learned Distributions

In color constancy (Chakrabarti, 2015), the model constructs a joint histogram $N[\hat{x}, y]$ from large-scale labeled data, followed by computation of $$L[\hat{x}, y] = \log [N[\hat{x}, y] / \sum_{\hat{x}'} N[\hat{x}', y]]$$. This log-likelihood serves as a lookup-based classifier, or can be parameterized and trained end-to-end to minimize an expected angular error on illumination estimates.

B. Latent-Space Augmentation via Embedding

In diffusion-based text-to-image generation, chromatic priors are projected or zero-padded to match CLIP token dimensionality, yielding $z_c$, which is prepended or concatenated to transformer input (Aggarwal et al., 2023). The transformer fuses chromatic and semantic signals through self-attention at every layer, thus integrating palette constraints during generation.

C. Cross-Attention Fusion with Compensated Images

For restoration tasks exhibiting strong color bias (e.g., underwater images), the chromatic-prior-modified image $\mathbf{y}$ is injected into a denoising U-Net at each denoising step via cross-attention (Shaahid et al., 15 Dec 2025). Learnable projections $Q, K, V$ extract features from both noisy latent $x_t$ and chromatic-prior image $\mathbf{y}$, producing attended features that steer each layer's computation.

D. Luminance and Semantic Conditioning during Colorization

Latent-space diffusion models concatenate luminance latents or semantic embeddings with the noisy sample at every step, typically via $1 \times 1$ convolutions and cross-attention blocks (Wang et al., 2024). Additional spatial priors from segmentation masks may be interpolated in the later denoising steps.

These approaches are unified by the principle that chromatic priors are not simply concatenated or injected at model input, but are fused repeatedly and adaptively—often at every inference or reconstruction step—through learned attention, projection, or classifier-based weighting.

3. Mathematical Formalization Across Domains

The precise mathematical realization of CPGC varies by application, but several key operational templates recur:

Color Constancy (Chakrabarti, 2015)

• For each candidate illuminant $i$, mapped chromaticities $\hat{x}_i(n)$ are obtained via $g(\mathbf{v}(n), \hat{m}_i)$.
• Per-pixel likelihoods $L[\hat{x}_i(n), y(n)]$ are summed to yield per-illuminant log-scores $l_i$:

$$l_i = \frac{\alpha}{N} \sum_{n=1}^N L[g(v(n), \hat{m}_i), y(n)] + \beta b_i$$

• The global illuminant is the expectation under the posterior $p_i \propto \exp(l_i)$.

Diffusion Prior (Text-to-Image, Colorization) (Aggarwal et al., 2023, Wang et al., 2024)

• The forward process is the usual DDPM:

$$q(z_t | z_{t-1}) = \mathcal{N}(z_t ; \sqrt{\alpha_t} z_{t-1}, (1-\alpha_t) I)$$

• The reverse process is conditioned:

$$p_\theta(z_{t-1} | z_t, z_c, c) = \mathcal{N}(z_{t-1} ; \mu_\theta(z_t, t, z_c, c), \sigma_t^2 I)$$

where $z_c$ can be a color histogram, luminance latent, or other chromatic prior.

• Noise-prediction loss is minimized, and during sampling, classifier-free guidance scales the chromatic prior's influence.

Cross-Attention in U-Net (Shaahid et al., 15 Dec 2025)

• At each relevant feature resolution:

$$A = \mathrm{Softmax}\left(\frac{ Q(x_t) K(\mathbf{y})^T }{ \sqrt{d_k} } \right), \qquad \mathrm{CA}(x_t, \mathbf{y}) = A V(\mathbf{y})$$

• This attended tensor is fused with U-Net features, with $\mathbf{y}$ being the chromatic-prior compensated image.

Decoder Alignment (Wang et al., 2024)

• Decoder skip connections fuse grayscale encoder features into color decoder via projection and additive fusion, ensuring chromatic consistency with the original structure and luminance.

4. Application Domains and Empirical Outcomes

CPGC finds application in several core vision tasks:

• Color Constancy: Accurate global illuminant estimation through per-pixel likelihood aggregation, outperforming contemporaneous methods (Chakrabarti, 2015).
• Text-to-Image Synthesis: Palette control without retraining large decoders, enabling prompt-independent chromatic consistency and improved semantic realism (Aggarwal et al., 2023). Quantitative metrics (Hellinger, KL, FID) demonstrate improvements over prior baselines.
• Underwater Image Enhancement: Mitigation of color cast and recovery of color fidelity across challenging conditions, outperforming traditional, CNN-, GAN-, and diffusion-based baselines in UCIQE/UIQM and qualitative structure preservation (Shaahid et al., 15 Dec 2025).
• Automatic Colorization: Saturated, semantically plausible color synthesis with fidelity to grayscale content and multimodal guidance (text, masks). The approach yields superior perceptual quality and user preference (Wang et al., 2024).

A summary table of major instantiations:

Domain	Chromatic Prior Form	Conditioning Mechanism
Color Constancy	(chrom, lumin) log-lhood $L$	Histogram lookup + aggregation
Text-to-Image Generation	L*a*b* histogram	Transformer token, attention
Underwater Enhancement	Lab-compensated RGB	Cross-attention U-Net
Colorization	Luminance latent, masks	Channel concat, cross-attn

5. Conditioning Mechanisms: Cross-Attention, Concatenation, and Aggregation

The fusion of chromatic priors is central to the CPGC paradigm and distinguishes effective schemes from baseline concatenation or static modulation:

• Cross-Attention: Aligns features from the prior and latent/noisy state through content-dependent weights, allowing dynamic spatially varying influence (Shaahid et al., 15 Dec 2025Aggarwal et al., 2023Wang et al., 2024).
• Channel Concatenation and Projection: Direct concatenation (e.g., grayscale latent, color-histogram vector) with subsequent $1 \times 1$ convolutional projection for learnable channel mixing (Wang et al., 2024).
• Statistical Aggregation: In classifier-based approaches, aggregation of per-pixel likelihoods or log-scores yields a consistent global estimate (Chakrabarti, 2015).

The cross-attention variants yield finer spatial and contextual control, crucial in applications where color distortion or spatially localized color guidance is necessary.

6. Training Protocols and Evaluation

Effective training of CPGC systems employs empirical priors as initialization, then optimizes end-to-end or with task-specific objectives:

• End-to-end Likelihood Tuning: Gradient-based optimization of $L[\hat{x}, y]$ for minimal angular error (Chakrabarti, 2015).
• Classifier-Free Guidance: Dropout of conditional priors during training to enable scalable sampling effects (Aggarwal et al., 2023Wang et al., 2024).
• Multi-term Losses: Enhanced losses incorporating pixel L1, perceptual (VGG), structure (SSIM), and frequency (FFT) terms (Shaahid et al., 15 Dec 2025).

Evaluation utilizes both standard reference metrics (FID, PSNR, colorfulness, user preference) and domain-specific ones (UCIQE, UIQM, histogram divergence). Empirical results consistently indicate superior palette control, structure preservation, and perceptual realism in CPGC-based systems across diverse vision tasks.

7. Implications, Advantages, and Empirical Observations

Chromatic prior-guided conditioning offers:

• Explicit palette or illuminant control not achievable through prompt engineering or unconditional models (Aggarwal et al., 2023).
• Improved realism, reduced color cast, and structure preservation due to spatially informed or semantically enhanced conditioning (Shaahid et al., 15 Dec 2025Wang et al., 2024).
• Empirical preference in human studies and strong quantitative superiority in reference metrics and no-reference image quality scores (Wang et al., 2024Shaahid et al., 15 Dec 2025).

This suggests that CPGC will remain foundational to future work in controlled image synthesis, restoration, and generative modeling where explicit color structure is crucial. Continued advances in conditioning mechanisms (cross-attention, dynamic priors, multimodal fusion) are likely, enabling even finer regulation of generative output and restoration quality.