BC-IHV: Conditioning the Color Space for Stable Rectified-Flow Low-Light Enhancement

Yi Ai; Zheng Chen; Yuanhao Cai; Yulun Zhang; Xiaokang Yang

cs.CVAug 22, 2026

BC-IHV: Conditioning the Color Space for Stable Rectified-Flow Low-Light Enhancement

Authors: Yi Ai, Zheng Chen, Yuanhao Cai, Yulun Zhang, Xiaokang Yang

Organizations: Shanghai Jiao Tong University · Johns Hopkins University

Abstract

Low-light image enhancement (LLIE) must correct ambiguous exposure without overwriting structure already supported by the input. Generative transport can model exposure ambiguity; however, its flexibility may also alter observable geometry and chromatic content. Moreover, fixed invertible color coordinates are usually treated only as representations, although their inverse mappings reshape the RGB-domain gradients received by the enhancement network. To address these issues, we propose Structure-Anchored Rectified Flow (SA-RF), which maintains correspondence through separate chromaticity/intensity stems, a scale-matched condition pyramid, and HybridAda. HybridAda assigns location-specific retrieval to spatial cross-attention and global exposure modulation to pooled AdaLN. We further introduce BC-IHV, a learnable Box--Cox polar color space whose analytically invertible intensity mapping controls the inverse-gradient dynamic range through a single exponent. This allows the representation to balance dark-range expansion and gradient conditioning instead of adopting a fixed linear or logarithmic law. Experiments on three LOL benchmarks, blind image-quality evaluation, and cross-dataset tests demonstrate consistent reconstruction and perceptual advantages over the sota. Controlled studies further support the effectiveness of both the proposed framework and color representation.

Figures & tables

Figure 1: Qualitative comparison with previous sota. On LOL-v2 Real, the proposed method shows stronger scene contrast and lower GT error under the shared error-map scale; darker values indicate smaller absolute RGB error.

Figure 2: Intensity-law diagnosis. (a) The mappings redistribute dark inputs. (b) Log-IHV creates a

1001{\times}

inverse-derivative range, whereas learned BC-IHV selects a

12{\times}

range. (c) LOL-v2 Real operating points: HVI and log-IHV form the matched endpoint probe; learned BC-IHV is paired with SA-RF. Together, they illustrate how the learned exponent selects an intermediate regime between dark-range expansion and gradient conditioning.

Figure 3: SA-RF with BC-IHV. (a) The low-light condition $c=\Phi_{\lambda}(y)$ is encoded by separate HV/I stems into $\{c_{0},c_{1},c_{2}\}$ , which anchor matched encoder/decoder scales. Tokens from $c_{2}$ provide spatial keys and values, while $g=f_{t}(t)+W_{p}\operatorname{GAP}(c_{2})$ controls global AdaLN. A 20-step Euler solver produces the enhanced latent before the closed-form inverse. (b) BC-IHV learns the Box–Cox exponent end-to-end. (c&d) Encoder/decoder residual blocks and HybridAda components: self-attention, cross-attention, MLP, and zero-initialized AdaLN. The condition pyramid maintains access to spatial evidence, while HybridAda separates local retrieval from global exposure control. Together, the two paths condition the transport on observed structure without fixing the enhanced exposure.

Figure 4: Optimization and inverse-gradient evidence on LOL-v2 Synthetic. (a) Strong-perceptual trajectories over the shared first 430 epochs; runs use matched hyperparameters and independent seeds. Thin lines are evaluations and thick lines are five-point means. (b) Exact inverse derivatives. (c) Brightness-bin mass and median derivative over 18M pixels from 900 training images. (d) Spatial derivative maps on a shared log scale. Shared axes and scales permit direct trajectory and spatial comparisons. The panels connect analytic derivative scaling to the observed optimization regime.

Method	Params (M)	NFE	FPS	LOL-v1					LOL-v2 Real					LOL-v2 Synthetic
				PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	BRISQ. $\downarrow$	C-IQA $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	BRISQ. $\downarrow$	C-IQA $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	BRISQ. $\downarrow$	C-IQA $\uparrow$
RetinexNet ( 34 )	0.44	1	423.0	18.92	0.427	0.470	51.79	0.493	16.10	0.401	0.543	56.39	0.542	17.14	0.761	0.255	27.44	0.571
KinD ( 45 )	8.02	1	256.2	23.01	0.843	0.156	31.08	0.370	23.77	0.874	0.121	30.42	0.400	18.32	0.795	0.252	27.30	0.376
Zero-DCE ( 8 )	0.08	1	1347.0	21.06	0.535	0.353	27.40	0.611	18.06	0.574	0.312	35.38	0.578	17.76	0.816	0.168	19.31	0.517
EnlightenGAN ( 13 )	8.64	1	345.5	21.24	0.643	0.331	10.72	0.430	18.64	0.675	0.309	18.54	0.426	16.57	0.775	0.212	16.71	0.488
KinD++ ( 44 )	8.27	1	53.3	24.16	0.842	0.155	29.30	0.409	22.21	0.843	0.175	34.08	0.397	19.26	0.806	0.226	28.05	0.447

Table 1: Unified LOL re-evaluation and efficiency. Best and second-best quality values are bold and underlined;

\dagger

marks the official multi-task checkpoint used for DarkIR.

Average over LOL-v1, LOL-v2 Real, and LOL-v2 Synthetic
Method	NIQE $\downarrow$	BRISQUE $\downarrow$	MUSIQ $\uparrow$	CLIP-IQA $\uparrow$	MANIQA $\uparrow$	TOPIQ-NR $\uparrow$	LOE $\downarrow$
Retinexformer	3.816	14.315	62.858	0.421	0.385	0.565	105.51
CIDNet	4.259	14.824	68.444	0.504	0.490	0.664	89.22
InterLight	4.158	15.603	68.645	0.459	0.465	0.647	97.62
SA-RF + BC-IHV (ours)	4.818	13.736	68.899	0.597	0.520	0.672	98.07
DICM

Table 2: Blind quality and cross-dataset transfer. Best/second-best values are bold/underlined.

Figure 5: Representative visual comparison. Five scenes from the LOL benchmarks compare representative method families. Cyan rectangles identify the same region enlarged inside every result; orange frames mark SA-RF+BC-IHV.

Intensity law: LOL-v2 Real
Representation	Model	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
HVI ( $\lambda=1$ )	SA-RF	22.54	0.847	0.131
Log-IHV	SA-RF	22.99	0.835	0.163
BC-IHV	SA-RF	24.21	0.857	0.097
Framework attribution: LOL-v2 Synthetic
HVI	CIDNet	25.13	0.939	0.045

Table 3: Controlled ablations of the intensity law and framework (left) and architecture components (right). Metrics are PSNR, SSIM, and LPIPS; block headers specify LOL-v2 Real or Synthetic.

Explore similar work

Retrieval-Augmented Generation-Based Color Restoration for Low-Light Image Enhancement
Aug 8, 2026Li-Wei Lu, Shaou-Gang MiaouLow-Light Image Enhancement Image Restoration
Department of Electronic Engineering, Chung Yuan Christian University, Taoyuan 320, Taiwan
Towards Color-Faithful Low-Light Image Enhancement via Adaptive Color Debiasing and Saturation Rectification
Aug 11, 2026Zhichen Yang, Rui Xu, Yuzhen Niu +3Image Enhancement Low-Light Image Enhancement
Fuzhou University, Fuzhou, Fujian, China
Beyond Illumination: A Conditional Mutual Information-Guided Network for Low-Light Image Enhancement
Aug 3, 2026Ya-nan Guan, Shaonan Zhang, Tao Dai +5Low-Light Image Enhancement
College of Artificial Intelligence, Nankai University, Tianjin, China · College of Computer Science and Software Engineering, Shenzhen University, China · Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China +2

Aug 8, 2026, eess.IVJ/K move · Enter open · S save

Retrieval-Augmented Generation-Based Color Restoration for Low-Light Image Enhancement

Li-Wei Lu, Shaou-Gang Miaou

Department of Electronic Engineering, Chung Yuan Christian University, Taoyuan 320, Taiwan

Low-Light Image Enhancement Image Restoration

Open paper AskPDF arXiv

Recent low-light image enhancement (LLIE) methods have driven brightness and structural fidelity close to that of normally-exposed images, yet their outputs still exhibit systematic color shifts such as greenish skies, yellowish faces, and warm-tinted white objects. We attribute this to end-to-end LLIE training coupling brightness, structure, and color within a single network, leaving the color channels weakly supervised. We recast color restoration as an independent sub-problem and decouple it from brightness enhancement, realizing it as a general-purpose post-processing module built on retrieval-augmented generation (RAG). Rather than relying solely on parametric color priors learned during training, the module dynamically retrieves a reference image from an external high-quality color knowledge base and injects its color distribution into a color-restoration network to correct residual bias. The design has three components: (i) a dual-index FAISS retriever built on intermediate VGG19 features, capturing textural and structural similarity through global mean and variance statistics; (ii) GlobalSPHistAdaIN, which reduces the reference spatial-preserving color histogram to a global color vector and modulates network features via adaptive instance normalization, removing dependence on pixel-level correspondence; and (iii) a residual formulation that predicts a color correction over the front-end output. Across LOLv1, LOLv2-Real, and LOLv2-Synthetic, the module consistently improves color-specific metrics, and it remains effective when the front end is swapped among CPGA-Net++, LLFormer, FLIGHTNet, and IAT, confirming cross-front-end generality. Ablations show that a VGG19 dual index outperforms CLIP-based retrieval, indicating that color restoration depends on textural and structural similarity rather than high-level semantics.

Abstract

Explore similar work

Retrieval-Augmented Generation-Based Color Restoration for Low-Light Image Enhancement

Towards Color-Faithful Low-Light Image Enhancement via Adaptive Color Debiasing and Saturation Rectification

Beyond Illumination: A Conditional Mutual Information-Guided Network for Low-Light Image Enhancement