Skip to content

LUCE: Constrained Curve-Domain Guidance for Training-Free Low-Light Enhancement with Hue-Preserving Decoupling

Conference: ECCV 2026
Paper: ECCV Official
Area: Image Restoration
Keywords: low-light image enhancement / diffusion model / training-free guidance / constrained curve domain / hue-preserving decoupling

TL;DR

Addressing the instability and chromatic distortions caused by competing pixel-space attribute gradients in training-free diffusion-based enhancement, LUCE reformulates sampler-side guidance as a constrained curve-domain luminance control problem with hue-preserving decoupling, driving a frozen DDPM sampler with a single scalar luminance energy.

Background & Motivation

Images captured under low-illumination environments suffer from severe degradations, including low visibility, noise amplification, and chromatic distortion, significantly deteriorating both human perceptual quality and downstream vision system performance. Classical approaches relying on histogram equalization or Retinex-based illumination estimation offer lightweight execution, yet frequently amplify sensor noise and introduce halo artifacts in complex, non-uniform low-light scenes. While supervised deep convolutional/transformer models and zero-reference curve-learning methods (e.g., Zero-DCE) deliver noticeable improvements, they remain constrained by labor-intensive paired dataset collection or generalization bottlenecks under extreme darkness.

Recently, generative diffusion models have demonstrated remarkable generative priors for recovering structural details and natural textures. Nonetheless, task-specific retraining or fine-tuning (e.g., DiffLL, LightenDiffusion) incurs prohibitive computational costs. An appealing alternative is the training-free paradigm (e.g., PGDiff, AGLLDiff), which adapts a frozen pretrained diffusion backbone at inference time by injecting hand-crafted pixel-space objectives for exposure, structure, and color. However, this line of work suffers from three inherent deficiencies: (i) gradient competition, where heterogeneous pixel-space loss terms pull sampling steps in conflicting directions, requiring tedious hyperparameter balancing and destabilizing the reverse trajectory; (ii) luminanceโ€“chroma entanglement, where optimizing unconstrained RGB intensities alters channel proportions, producing unnatural saturation and chromatic shifts; and (iii) implicit brightness evolution, which lacks an explicit global model to enforce a smooth, monotonic exposure trajectory.

To fundamentally resolve multi-objective gradient conflict and color deviation, this paper reframes sampler-side enhancement as a constrained luminance control problem within a low-dimensional monotonic curve space. Core idea: project luminance guidance into a single scalar energy defined over a low-dimensional monotonic lookup table to eliminate gradient competition, while executing luminance edits via per-pixel scalar gains that preserve RGB vector directions to guarantee rigorous hue preservation.

Method

Overall Architecture

LUCE operates entirely on top of a frozen pretrained DDPM diffusion backbone, injecting gradient corrections exclusively during the final \(K\) reverse sampling steps (default \(K=10\)). Given a low-light input \(x_{\text{in}}\), at denoising step \(t\), the backbone UNet produces an estimate of the clean image \(\hat{x}_0(x_t, t)\), from which the linear-domain luminance map \(Y_t\) is computed. The overall method connects three core components sequentially: 1. Unified Luminance Target (UM*): Formulates an explicit global target \(Y_t^\star\) by taking a time-dependent convex combination of the current predicted luminance \(Y_t\) and an input-derived exposure prior \(E(x_{\text{in}})\); 2. Curve-Guided Luminance Energy (CMM*): Fits a monotonic lookup table (LUT) curve \(C_{\phi_t}\) in the 1D luminance domain and establishes a single scalar energy \(E_t(x_t) = \mathcal{L}_{\text{curve}}\), which directs the DDPM sampler update through a single coherent gradient; 3. Hue-Preserving Chromatic Decoupling (CDM*): Computes per-pixel scalar gains based on the curve-mapped luminance and scales the RGB color vectors uniformly, preserving hue and preventing unnatural saturation overshoots.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Low-light input x_in and noisy state x_t"] --> B["Frozen DDPM UNet denoising prediction<br/>estimate clean image x^0_t"]
    B --> C["Unified Luminance Target<br/>fuse predicted luminance Y_t with exposure prior E(x_in) into Y*_t"]
    C --> D["Curve-Guided Luminance Energy<br/>fit monotonic LUT curve and evaluate single scalar energy E_t(x_t)"]
    D --> E["Hue-Preserving Chromatic Decoupling<br/>compute per-pixel scalar gain s_t to scale RGB vectors preserving hue"]
    E --> F["Inject single-energy gradient into sampler<br/>execute discrete DDPM update to x_t-1"]

Key Designs

1. Unified Luminance Target: Dynamically blending denoising predictions with input exposure priors

Existing training-free approaches combine multiple pixel-space constraints (e.g., exposure alignment, well-exposed region preservation, mean brightness matching), inevitably inducing gradient interference. LUCE resolves this issue by consolidating all luminance objectives into a single reference distribution \(Y_t^\star\) prior to sampler guidance. The system first extracts input luminance \(Y_{\text{in}}\) in linear RGB via \(Y = 0.299R + 0.587G + 0.114B\), applies global gamma correction \(\hat{Y}_\gamma = Y_{\text{in}}^\gamma\) (with \(\gamma=0.6\)), and performs affine rescaling with clamping to match a target mean luminance \(\mu_{\text{tar}}=0.5\), constructing a reliable exposure prior \(E(x_{\text{in}})\). At denoising step \(t\), the current predicted luminance \(Y_t\) and \(E(x_{\text{in}})\) are blended element-wise: $\(Y_t^\star = \beta_t Y_t + (1 - \beta_t) E(x_{\text{in}})\)$ The schedule \(\beta_t \in [0.3, 0.9]\) smoothly arbitrates between generative fidelity and exposure correction: earlier guided steps use smaller \(\beta_t\) to lift underexposed regions quickly, while as \(t \to 0\), \(\beta_t\) approaches 1.0 to anchor sampling to the diffusion model's high-frequency textural synthesis.

2. Curve-Guided Luminance Energy: Low-dimensional monotonic LUT parameterization with shape regularizers

Directly penalizing pixel-wise differences \(|Y_t - Y_t^\star|\) permits uncoordinated local fluctuations, often degrading global exposure consistency. LUCE instead restricts luminance adjustments to a global 1D monotonic curve transformation. The curve \(C_\phi: [0, 1] \to [0, 1]\) is parameterized as a piecewise-linear lookup table with \(B=32\) uniform knots and knot values \(\phi = \{\phi_0, \dots, \phi_B\}\), strictly enforcing monotonicity \(0 \le \phi_0 \le \phi_1 \le \dots \le \phi_B \le 1\). To preserve natural contrast, the curve optimization objective incorporates an \(\ell_1\) second-order difference smoothness penalty \(\mathcal{L}_{\text{smooth}}(\phi)\) and an \(\ell_2\) identity-mapping shape penalty \(\mathcal{L}_{\text{shape}}(\phi)\): $\(\mathcal{L}_{\text{curve}}(\phi) = \frac{1}{N} \sum_{p=1}^N \big| C_\phi(Y_t(p)) - Y_t^\star(p) \big| + \lambda_1 \mathcal{L}_{\text{smooth}}(\phi) + \lambda_2 \mathcal{L}_{\text{shape}}(\phi)\)$ The optimal curve parameters \(\phi_t\) are obtained via a few projected gradient steps in low-dimensional space (\(B \ll N\)). When evaluating \(\nabla_{x_t} E_t(x_t)\) for sampler guidance, \(\phi_t\) is held constant, reducing the guidance vector strictly to a single luminance-matching gradient. Because smoothness and shape regularizers are completely resolved inside curve space, no competing gradient terms are backpropagated to the diffusion state.

3. Hue-Preserving Chromatic Decoupling: Distortion-free scalar gain modulation

Once the curve-adjusted luminance \(\tilde{Y}_t(p) = C_{\phi_t}(Y_t(p))\) is computed, realizing this change across RGB channels without color bias is paramount. Independent channel adjustments warp the spatial orientation of the RGB color vector, resulting in hue drift and unnatural chromatic saturation. LUCE exploits the geometric principle that uniform positive scaling of an RGB vector strictly preserves its chromatic direction and hue before gamut clipping. The per-pixel scalar gain is defined as: $\(s_t(p) = \frac{\tilde{Y}_t(p)}{\max(Y_t(p), \varepsilon)}\)$ and the modulated color vector is evaluated as \(\tilde{c}_t(p) = \text{clip}_{[0.5, 3.0]}(s_t(p)) \cdot c_t(p)\). In the absence of clipping, the transformation satisfies \(w^\top \tilde{c}_t(p) = \tilde{Y}_t(p)\) where \(w = [0.299, 0.587, 0.114]^\top\), completely separating luminance scaling from chromatic attributes and preventing over-saturation or greenish/purplish color casts.

Loss & Training

LUCE is a fully training-free framework; the pretrained diffusion backbone weights remain completely frozen during inference. The sampler incorporates gradient guidance into the standard discrete DDPM update over the final \(K=10\) steps: $\(x_{t-1} = \text{DDPMUpdate}_\theta(x_t, t) - \lambda_t \nabla_{x_t} E_t(x_t)\)$ where the time-dependent guidance strength \(\lambda_t\) ramps smoothly from 0 to 1.0. Hyperparameters are fixed across all benchmarks: \(B=32\), \(\lambda_1 = 0.01\), \(\lambda_2 = 0.1\), and scalar gains clipped to \([0.5, 3.0]\). All evaluations execute on a single NVIDIA A100 GPU without dataset-specific tuning.

Key Experimental Results

Main Results

On paired benchmarks LOL-v1, LOL-v2 (combined Real+Synthetic), and LSRW, LUCE is evaluated against representative supervised (S) and unsupervised (U) methods (Table 1). LUCE consistently leads the unsupervised category, obtaining superior structural (SSIM) and perceptual (LPIPS) metrics.

Type Method LOL-v1 (PSNRโ†‘/SSIMโ†‘/LPIPSโ†“) LOL-v2 (R+S) (PSNRโ†‘/SSIMโ†‘/LPIPSโ†“) LSRW (PSNRโ†‘/SSIMโ†‘/LPIPSโ†“)
S Retinex-Net 17.62 / 0.64 / 0.38 16.62 / 0.58 / 0.40 15.58 / 0.41 / 0.39
S KinD 21.02 / 0.84 / 0.20 17.93 / 0.73 / 0.31 15.51 / 0.38 / 0.39
S URetinex-Net 19.84 / 0.82 / 0.13 20.56 / 0.86 / 0.24 18.27 / 0.52 / 0.30
S DiffLL 27.93 / 0.88 / 0.11 25.47 / 0.89 / 0.11 19.27 / 0.55 / 0.30
U Zero-DCE 14.86 / 0.55 / 0.34 16.52 / 0.68 / 0.26 15.84 / 0.45 / 0.31
U RUAS 16.44 / 0.49 / 0.27 15.32 / 0.58 / 0.49 14.31 / 0.48 / 0.47
U SCI 14.78 / 0.53 / 0.33 13.38 / 0.61 / 0.33 15.24 / 0.42 / 0.45
U NeRCo 19.81 / 0.74 / 0.24 17.19 / 0.64 / 0.39 18.82 / 0.51 / 0.34
U LightenDiffusion 20.44 / 0.81 / 0.19 19.24 / 0.83 / 0.17 18.54 / 0.54 / 0.31
U AGLLDiff 21.81 / 0.84 / 0.16 19.59 / 0.83 / 0.21 16.92 / 0.56 / 0.32
U LUCE (Ours) 22.81 / 0.85 / 0.15 20.21 / 0.84 / 0.19 18.80 / 0.56 / 0.32

On unpaired benchmarks (LIME, NPE, MEF, DICM, VV), perceptual quality (BRISQUE), residual noise (NL), and computational efficiency are benchmarked in Table 2. Compared with the training-free diffusion baseline AGLLDiff, LUCE cuts inference time by nearly half (from 20.43s to 10.53s) while achieving lower perceptual degradation and noise.

Type Method Params (M) MACs (G) Time (s) BRISQUEโ†“ NLโ†“
S Retinex-Net 0.555 87.28 0.093 27.10 3.25
S DiffLL 22.08 21.88 0.340 24.13 0.68
U Zero-DCE 0.079 20.92 0.016 21.76 1.57
U SCI 0.00026 0.095 0.002 16.73 0.85
U NeRCo 23.30 822.8 0.949 22.81 0.65
U LightenDiffusion 26.54 2257.42 9.539 18.53 1.11
U AGLLDiff 552.81 1114.68 20.43 19.55 0.80
U LUCE (Ours) 556.22 1106.35 10.53 18.03 0.76

Ablation Study

To evaluate the contribution of each individual module (UM, CMM, and CDM*), systematic ablation experiments were conducted on LOL-v2 and the five unpaired benchmarks (Table 3).

Config LOL-v2 PSNRโ†‘ LOL-v2 SSIMโ†‘ LOL-v2 LPIPSโ†“ Unpaired BRISQUEโ†“ Unpaired NLโ†“ Note
w/o UM* (remove unified target) 19.70 0.832 0.205 20.71 1.45 Lacks exposure reference; largest drops in PSNR/SSIM and severe noise
w/o CMM* (replace curve with pixel โ„“1) 19.85 0.836 0.200 19.86 1.39 Uncoordinated pixel drift causes flat contrast and washed-out surfaces
w/o CDM* (remove chromatic decoupling) 19.78 0.834 0.208 19.03 1.08 Unbalanced channel scaling induces noticeable hue shifts and higher LPIPS
LUCE (full model) 20.21 0.840 0.192 18.03 0.76 Synergistic combination achieves optimal exposure, detail, and chromatic fidelity

Key Findings

  • The unified target (UM*) anchors global brightness trajectory: Removing UM* causes PSNR on LOL-v2 to fall from 20.21 dB to 19.70 dB, and residual noise NL nearly doubles from 0.76 to 1.45. This confirms that an explicit exposure reference prevents sampling drift under low visibility.
  • Monotonic curve constraints (CMM*) prevent uncoordinated pixel artifacts: Guiding with raw pixel-wise โ„“1 loss degrades perceptual quality and causes washed-out textures. Constraining updates to a 32-knot monotonic LUT coordinates global tone progression while cutting backpropagation overhead.
  • Chromatic decoupling (CDM*) is essential for color fidelity: Discarding CDM* causes the sharpest deterioration in LPIPS (0.192 \(\to\) 0.208) and introduces severe color casts in shadow regions, verifying the necessity of maintaining RGB vector direction during brightness enhancement.

Highlights & Insights

  • Reconciling multi-objective pixel guidance into low-dimensional curve control: Rather than juggling conflicting pixel-space penalties for exposure, contrast, and color, LUCE absorbs regularizations inside a 1D monotonic curve representation, allowing a single scalar energy gradient to guide diffusion steps harmoniously.
  • Mathematically grounded hue-preserving scalar gain: Scaling RGB color vectors uniformly by a positive scalar gain provides an elegant mechanism to modify luminance while preserving chromatic direction, fundamentally avoiding color casts without complex color-space conversions.
  • Practical efficiency in training-free diffusion restoration: By optimizing curve parameters in low-dimensional space and restricting guidance to the final 10 denoising steps, LUCE cuts runtime in half compared to AGLLDiff, setting an effective precedent for lightweight diffusion guidance.

Limitations & Future Work

  • Hand-crafted heuristic exposure prior: The current exposure prior relies on fixed global gamma correction (\(\gamma=0.6\)) and a constant target mean (\(\mu_{\text{tar}}=0.5\)), which may struggle with complex high-dynamic-range (HDR) scenes containing both harsh spot lighting and deep shadows.
  • Inference latency versus lightweight CNNs: While 10.53s per 256ร—256 image is competitive among diffusion-based restoration methods, it remains considerably slower than real-time curve CNNs (e.g., Zero-DCE's tens of milliseconds).
  • Future directions: Developing content-adaptive exposure priors, extending curve-domain guidance to latent diffusion architectures (e.g., Stable Diffusion / Flux), and exploring temporal coherence for training-free low-light video enhancement.
  • vs Zero-DCE [6]: Zero-DCE trains lightweight CNNs to estimate pixel-wise curve parameters offline; while fast, it struggles in extreme darkness due to lack of generative priors. LUCE is fully training-free, exploiting pretrained diffusion models to synthesize rich textures with a globally regularized 1D LUT.
  • vs AGLLDiff [24]: AGLLDiff guides frozen diffusion via multi-attribute pixel losses with heuristic weights, often suffering from gradient conflicts and color shifts. LUCE introduces curve-domain single-energy guidance with chromatic decoupling, improving PSNR/SSIM, eliminating color casts, and cutting runtime by 50%.
  • vs LightenDiffusion [15]: LightenDiffusion requires task-specific training of an unsupervised latent-Retinex diffusion model. LUCE requires zero training or architectural modifications, serving as an off-the-shelf plug-and-play guidance framework.

Rating

  • Novelty: โญโญโญโญโ˜† [Reformulates diffusion guidance as low-dimensional monotonic curve control with hue-preserving decoupling, elegantly resolving gradient conflict]
  • Experimental Thoroughness: โญโญโญโญโญ [Evaluated across 3 paired and 5 unpaired benchmarks with rigorous distortion, perceptual, noise, and efficiency metrics]
  • Writing Quality: โญโญโญโญโญ [Clear mathematical formulations, coherent motivation, and precise ablation analysis]
  • Value: โญโญโญโญโ˜† [Offers an insightful, practical paradigm for controllable training-free diffusion sampling in low-level restoration]