Region-Constrained Color Preservation

• Region-constrained color preservation is a technique that maintains color fidelity by applying strict spatial or parametric constraints to designated image regions.
• It employs methods like multi-phase constrained optimization and diffusion-based ROI guidance to balance perceptual similarity with accessibility, achieving low ΔE values and WCAG compliance.
• This approach supports applications in accessible web design, image compositing, and generative modeling by decoupling color preservation from other image modifications to maintain brand integrity and visual realism.

Region-constrained color preservation methods are a class of algorithms and frameworks that enforce color consistency and fidelity in specific spatial or parametric regions of an image, model output, or color palette. This is achieved by applying explicit region-based constraints—typically on hue, saturation, or entire color profiles—while optimizing other degrees of freedom to meet task-specific criteria, such as contrast accessibility, image harmonization, or generative model guidance. These techniques are foundational for applications in accessible web design, image compositing, and diffusion-based image generation, where strict region-based color adherence is essential to visual coherence, perceptual realism, or brand integrity.

1. Problem Domains and Definitions

Region-constrained color preservation encompasses problems where color transformations must be spatially or parametrically restricted to maintain critical semantic or aesthetic properties. Two main domains have emerged:

• Parametric region constraints: In palette or single-color optimization, the method restricts color change to specific subspaces (e.g., fixing hue $H$ in the OKLCH cylinder), allowing lightness $L$ and chroma $C$ to vary under task constraints such as contrast or perceptual indistinguishability (R, 4 Dec 2025).
• Spatial region constraints: In images, preservation targets specific regions-of-interest (ROIs) delineated by a binary mask, facilitating local color control within semantic or user-defined areas (e.g., logo, product area, or pasted object), while background or complementary regions are separately handled (Ahuja et al., 24 Jan 2026, Zhu et al., 2022).

The fundamental objective is to match or constrain the color properties of designated regions (either in color space or spatially) while optimizing for higher-level perceptual, accessibility, or task-driven goals.

2. Mathematical Formulations

2.1 Parametric Hue-Region Constraint (OKLCH Domain, Accessibility Context)

Given an original color $c_0 = (L_0, C_0, H_0)$ in OKLCH, and a background $c_{\mathrm{bg}}$, the target is to adjust $c_0$ to $c' = (L', C', H')$ such that:

• $H' = H_0$ (region constraint: hue is preserved exactly)
• $\mathrm{contrast}(c', c_{\mathrm{bg}}) \ge \tau$ (WCAG, with $\tau=4.5$ or $7.0$)
• $c'$ is as perceptually close to $c_0$ as possible, measured by $\Delta E_{2000}(c_0, c')$
• $c'$ remains sRGB gamut valid

Mathematically: $$\begin{aligned} \min_{(L', C', H')} \quad & \Delta E_{2000}(c_0, c') \ \text{subject to} \quad & H' = H_0 \ & \mathrm{contrast}(\mathrm{OKLCH\to sRGB}(L', C', H'),\,c_{\mathrm{bg}}) \ge \tau \ & 0 \leq L' \leq 1; \quad 0 \leq C' \leq C_{\max} \ & \mathrm{OKLCH\to sRGB}(L', C', H') \in [0, 1]^3 \ \end{aligned}$$ (R, 4 Dec 2025)

2.2 Spatial Region Constraint (Diffusion/Compositing Contexts)

Given an image, a binary ROI mask $M\in\{0,1\}^{H\times W}$, and a target color or local statistics, the method enforces color preservation/guidance inside the ROI while background is kept fixed or adaptively harmonized. The objective may use composite losses defined on pixelwise or distributional measures between generated ROI colors and targets, e.g., mean, CVaR-tail, or log-sum-exp penalties in CIE Lab or RGB space (Ahuja et al., 24 Jan 2026, Zhu et al., 2022).

3. Algorithmic Approaches

3.1 Multi-Phase Constrained Optimization (Palette/Color Pair Setting)

The approach introduced in "Perceptually-Minimal Color Optimization for Web Accessibility: A Multi-Phase Constrained Approach" (R, 4 Dec 2025) employs a three-phase solver:

1. Binary Search on Lightness $L$: With $H$ and $C$ fixed, binary search adjusts $L$ toward the necessary contrast, respecting an upper bound on $\Delta E$.
2. Gradient Descent on $(L, C)$: Both $L$ and $C$ are optimized under a penalized cost combining WCAG constraint violation, perceptual threshold excess, and the raw $\Delta E$.
3. Progressive Constraint Relaxation: A sweep over increasing perceptual budgets $\delta_E^{\max}$, iterating previous phases to find the minimum $\Delta E$ solution providing the required contrast.

A high-level pseudocode sketch is as follows:

GenerateAccessibleColor(c₀, c_bg, large_text): If contrast(c₀, c_bg) ≥ τ_target: return c₀ For δₑᵐᵃˣ in Θ: c₁ ← BinarySearchLightness(c₀, c_bg, δₑᵐᵃˣ, τ_target) c₂ ← GradientDescentOKLCH(c₀, c_bg, δₑᵐᵃˣ, τ_target) Evaluate both; keep best feasible by ΔE and contrast Return best c′ (or c₀ if none)

This method yields median perceptual change $\Delta E_{2000} = 0.76$ for successful pairs, with $88.51\%$ being imperceptible ($\Delta E_{2000} < 2$), and a runtime of $0.876$ ms per pair (R, 4 Dec 2025).

3.2 Inference-Time Loss-Guided ROI Preservation in Diffusion Sampling

The approach described in (Ahuja et al., 24 Jan 2026) integrates three major mechanisms into Stable Diffusion:

1. ROI-Based Inpainting: Input masking and latent-space separation restrict generation to the target ROI while providing spatial selectivity.
2. Background-Latent Re-Imposition: At each denoising step, the latent background outside the ROI is restored from the original canvas, eliminating color drift.
3. Latent Nudging via Composite Loss: A composite loss in CIE Lab addresses mean, CVaR-tail, and soft-maximum (LSE) of color errors over the ROI. Where guidance is active, gradients are backpropagated to the latent only in the ROI, steering generation to match the target color distribution.

Guidance is suppressed during early noisy steps and increases in later steps per a time-dependent “late-start gate” schedule.

Table: Principal Loss Components in (Ahuja et al., 24 Jan 2026)

Loss Term	Targeted Statistic	Purpose
$\ell_\mathrm{mean}$	Region mean color	Controls global shift
$\ell_\mathrm{tail}$	CVaR-tail (worst-case pixels)	Suppresses local error spikes
$\ell_\mathrm{max}$	Log-sum-exp soft max	Caps outlier deviation

The full algorithm is training-free, integrating with standard inpainting pipelines.

3.3 Regional Feature Matching in Harmonization

In image harmonization, region-constrained preservation is accomplished via content-appearance affinity between masked foreground and reference background regions (Zhu et al., 2022):

• At high spatial resolution (PTL module), similarity matrices determine how each foreground location adapts color statistics (mean, variance) from matched local background regions.
• At deeper layers (LTL module), self-attention aggregates coarse-level location dependencies, refining foreground adaptation.
• Residual decoding ensures that important foreground image detail is preserved.

The harmonization system is trained exclusively with a foreground-weighted MSE loss.

4. Quantitative Results and Performance

• Palette Region-Constrained Optimization: On 10,000 color pairs, the algorithm of (R, 4 Dec 2025) achieved $77.22\%$ success in obtaining WCAG-compliant accessible pairs, with $88.51\%$ of these yielding imperceptible change ($\Delta E_{2000} < 2.0$). The median time per pair is $0.876$ ms.
• Diffusion ROI Preservation: Without guidance, mean ROI $\Delta E_{2000}$ is $55.2$; composite-loss guidance reduces it to $29.2$ with most pixels receiving $\Delta E_{2000}<10$. Mean-only methods yield visually salient local color failures, while CVaR-guided outputs demonstrate spatially uniform and on-target hues (Ahuja et al., 24 Jan 2026).
• Harmonization via Region Reference: The method of (Zhu et al., 2022) surpasses prior art on the iHarmony4 benchmark, with PSNR $37.97$ dB and foreground MSE $282.69$, showing consistent improvements across all data subsets.

5. Implementation Considerations

• Complexity and Speed: The multi-phase OKLCH approach is highly efficient; the theoretically worst-case scenario involves up to $2,040$ evaluations but, due to early stopping, median runtime is sub-millisecond (R, 4 Dec 2025).
• Guidance Timing: In diffusion-based methods (Ahuja et al., 24 Jan 2026), applying color guidance only after early denoising stages is critical to stability.
• ROI Mask Sizing: In both (Ahuja et al., 24 Jan 2026) and (Zhu et al., 2022), the binary mask defining the preservation region is up/downsampled as needed for compatibility with latent or feature map resolution.

6. Applications and Brand/Design Considerations

Constrained region color preservation supports:

• Accessibility and Brand Consistency: Fixing the hue region ensures that palette adjustments for WCAG accessibility do not compromise brand hues (R, 4 Dec 2025).
• Spatially Localized Editing: Enforces strict color targets in generative inpainting, product visualization, and fashion workflows, enabling precise control over part of the image without affecting other regions (Ahuja et al., 24 Jan 2026).
• Visual Realism and Image Compositing: By adaptively matching background appearance to foreground regions in compositing tasks, harmonization achieves consistent lighting, chromaticity, and realistic integration (Zhu et al., 2022).

The region-constrained approach generalizes to any setting where certain color DOFs (e.g., hue, temperature) or spatial regions are fixed due to semantic or design constraints.

7. Broader Impact and Extensions

Region-constrained color preservation constitutes a modular framework applicable across accessibility, generative modeling, visual harmonization, and palette management domains. Its core principle—minimizing a perceptual or statistical distance within algorithmically defined bounds in perceptually uniform spaces—enables practitioners to decouple preservation and adaptation along configurable axes, supporting robust task constraints without sacrificing critical design intent or visual coherence (R, 4 Dec 2025, Ahuja et al., 24 Jan 2026, Zhu et al., 2022).

A plausible implication is the extensibility to domains requiring compliance with perceptual metrics (e.g., color-blindness simulation, palette expansion, neural image editing) where constraints must be enforced in both parametric and spatial regions for reliable, automated control.