HVI-CIDNet+: Advanced Color–Intensity Decoupling

• The paper introduces HVI-CIDNet+, a framework that combines the novel HVI color space with a dual-branch decoupling architecture to restore chromaticity and intensity in extremely dark images.
• It employs nested U-shaped subnetworks with cross-attentive modules and vision-language priors to effectively address color distortion and brightness restoration in low-light conditions.
• Empirical tests on standard LLIE benchmarks demonstrate improved PSNR, SSIM, and LPIPS metrics over previous CIDNet models, validating the benefits of prior-guided and region-adaptive refinement.

Searching arXiv for the primary HVI-CIDNet+ paper and closely related HVI/CIDNet work to ground the article in current literature. Searching arXiv for "HVI-CIDNet+" and "HVI: A New Color Space for Low-light Image Enhancement". Color and Intensity Decoupling Network+ (HVI-CIDNet+) is a low-light image enhancement framework that combines the Horizontal/Vertical-Intensity color space with a dual-branch decoupling architecture for chromaticity and intensity restoration. In the formulation introduced in “HVI-CIDNet+: Beyond Extreme Darkness for Low-Light Image Enhancement,” the model is built upon the HVI color space to restore damaged content and mitigate color distortion in extremely dark regions, while the underlying HVI representation itself was introduced as a new color space for LLIE, defined by polarized HS maps and learnable intensity (Yan et al., 9 Jul 2025, Yan et al., 27 Feb 2025).

1. Origins and problem formulation

HVI-CIDNet+ emerged from a line of LLIE work arguing that standard sRGB processing is intrinsically unstable under severe underexposure because brightness and color are strongly coupled across the three RGB channels. The same literature also identified a second failure mode in HSV-based enhancement: although HSV decouples brightness and color, it introduces significant red and black noise artifacts. The HVI color space was proposed to address these issues, and CIDNet was introduced as a Color and Intensity Decoupling Network operating in that space (Yan et al., 27 Feb 2025). The later extension explicitly named “Color and Intensity Decoupling Network+” adds vision-language priors and region-adaptive refinement to address extreme darkness, where semantic content and reliable chromatic cues are often missing (Yan et al., 9 Jul 2025).

A recurrent misconception in the surrounding literature is that the original HVI/CIDNet papers already define a distinct “plus” architecture. The original HVI work defines HVI and CIDNet, and discusses possible extensions, but does not formally define a separate HVI-CIDNet+ model; the explicit “Color and Intensity Decoupling Network+” designation appears in the later work focused on “Beyond Extreme Darkness” (Yan et al., 27 Feb 2025, Yan et al., 9 Jul 2025).

The motivating pathologies are specific. In sRGB, enhancement amplifies both signal and noise while preserving the entanglement of chromaticity and brightness. In HSV, the hue axis is discontinuous around red, and the mapping near the black plane is unstable. HVI-CIDNet+ retains the decoupling advantage of HSV-like parameterizations while replacing the problematic hue geometry and low-intensity behavior with a continuous HV chromatic plane and a learnable intensity collapse mechanism (Yan et al., 9 Jul 2025).

2. HVI color space

The HVI representation starts from the Max-RGB intensity map

$$\mathbf{I}_{\max}(x) = \max_{\mathbf{c} \in \{R,G,B\}} \big(\mathbf{I}_c(x)\big),$$

which plays the role of the intensity channel. Saturation follows the HSV definition,

$$\mathbf{S} = \begin{cases} 0, & \mathbf{I}_{\max} = \mathbf{0},\ \dfrac{\Delta}{\mathbf{I}_{\max}}, & \mathbf{I}_{\max} \neq \mathbf{0}, \end{cases}$$

with $\Delta = \mathbf{I}_{\max} - \min(\mathbf{I}_c)$, and hue is inherited from the standard piecewise HSV formula (Yan et al., 27 Feb 2025).

The key change is the replacement of the scalar hue axis by a continuous orthogonal embedding. In the simplified formulation used for explaining HVI-CIDNet+, hue is polarized as

$$h = \cos\left(\frac{\pi \mathbf{h}}{3}\right), \qquad v = \sin\left(\frac{\pi \mathbf{h}}{3}\right).$$

This maps hue to a 2D plane, so red at the wrap-around boundary is no longer numerically discontinuous. In the more general HVI formulation, the original paper also defines a color-perceptual map $\mathbf{P}_\gamma$ and a function-density term $T(\cdot)$, allowing $\gamma_G$, $\gamma_B$, and $T$ to modulate hue sensitivity and color-plane density (Yan et al., 27 Feb 2025).

Low-intensity instability is handled by the learnable intensity collapse

$$\mathbf{C}_k = \sqrt[k]{ \sin\left( \frac{\mathbf{I}_{\max}\pi}{2} \right) + \varepsilon },$$

with $$\mathbf{S} = \begin{cases} 0, & \mathbf{I}_{\max} = \mathbf{0},\ \dfrac{\Delta}{\mathbf{I}_{\max}}, & \mathbf{I}_{\max} \neq \mathbf{0}, \end{cases}$$0 and $$\mathbf{S} = \begin{cases} 0, & \mathbf{I}_{\max} = \mathbf{0},\ \dfrac{\Delta}{\mathbf{I}_{\max}}, & \mathbf{I}_{\max} \neq \mathbf{0}, \end{cases}$$1. The final HV components are then

$$\mathbf{S} = \begin{cases} 0, & \mathbf{I}_{\max} = \mathbf{0},\ \dfrac{\Delta}{\mathbf{I}_{\max}}, & \mathbf{I}_{\max} \neq \mathbf{0}, \end{cases}$$2

where $$\mathbf{S} = \begin{cases} 0, & \mathbf{I}_{\max} = \mathbf{0},\ \dfrac{\Delta}{\mathbf{I}_{\max}}, & \mathbf{I}_{\max} \neq \mathbf{0}, \end{cases}$$3 in the general formulation, and $$\mathbf{S} = \begin{cases} 0, & \mathbf{I}_{\max} = \mathbf{0},\ \dfrac{\Delta}{\mathbf{I}_{\max}}, & \mathbf{I}_{\max} \neq \mathbf{0}, \end{cases}$$4 in the simplified setting adopted in the reported experiments (Yan et al., 27 Feb 2025). The resulting HVI tensor is therefore a three-channel representation $$\mathbf{S} = \begin{cases} 0, & \mathbf{I}_{\max} = \mathbf{0},\ \dfrac{\Delta}{\mathbf{I}_{\max}}, & \mathbf{I}_{\max} \neq \mathbf{0}, \end{cases}$$5.

The inverse mapping, termed PHVIT, reconstructs HSV and then sRGB. Given enhanced HVI channels $$\mathbf{S} = \begin{cases} 0, & \mathbf{I}_{\max} = \mathbf{0},\ \dfrac{\Delta}{\mathbf{I}_{\max}}, & \mathbf{I}_{\max} \neq \mathbf{0}, \end{cases}$$6, the inverse first divides out $$\mathbf{S} = \begin{cases} 0, & \mathbf{I}_{\max} = \mathbf{0},\ \dfrac{\Delta}{\mathbf{I}_{\max}}, & \mathbf{I}_{\max} \neq \mathbf{0}, \end{cases}$$7 and then computes

$$\mathbf{S} = \begin{cases} 0, & \mathbf{I}_{\max} = \mathbf{0},\ \dfrac{\Delta}{\mathbf{I}_{\max}}, & \mathbf{I}_{\max} \neq \mathbf{0}, \end{cases}$$8

before applying the standard HSV-to-sRGB transform (Yan et al., 2024). This representation is used not only as a preprocessing transform but also as a supervised output space.

3. Dual-branch CIDNet+ architecture

HVI-CIDNet+ uses two nested U-shaped subnetworks, one for the HV chromatic channels and one for the I intensity channel. The input low-light image is first transformed to HVI, yielding $$\mathbf{S} = \begin{cases} 0, & \mathbf{I}_{\max} = \mathbf{0},\ \dfrac{\Delta}{\mathbf{I}_{\max}}, & \mathbf{I}_{\max} \neq \mathbf{0}, \end{cases}$$9 and $\Delta = \mathbf{I}_{\max} - \min(\mathbf{I}_c)$0. In parallel, a pre-trained DA-CLIP encoder extracts two global priors from the same low-light image: latent semantic priors $\Delta = \mathbf{I}_{\max} - \min(\mathbf{I}_c)$1 and degraded representations $\Delta = \mathbf{I}_{\max} - \min(\mathbf{I}_c)$2 (Yan et al., 9 Jul 2025).

The core network remains a color–intensity decoupling architecture. The HV branch is dedicated to chromatic processing, color correction, and denoising. The I branch is dedicated to brightness mapping and content restoration. These branches are not independent: they exchange features through cross-attentive interaction blocks. In the original CIDNet formulation, this interaction is implemented by Lightweight Cross-Attention modules, which combine a Cross Attention Block with an Intensity Enhance Layer in the I branch and a Color Denoise Layer in the HV branch (Yan et al., 2024). HVI-CIDNet+ retains this decoupled logic but augments it with prior-guided fusion and region-adaptive refinement for extreme darkness (Yan et al., 9 Jul 2025).

The architectural motivation is explicit. In extremely dark regions, local evidence may be insufficient for either color restoration or content reconstruction. HVI-CIDNet+ therefore leverages abundant contextual and degraded knowledge extracted from low-light images using pre-trained vision-LLMs, integrated via a novel Prior-guided Attention Block. Within the PAB, latent semantic priors can promote content restoration, while degraded representations guide precise color correction, both particularly in extremely dark regions through the meticulously designed cross-attention fusion mechanism (Yan et al., 9 Jul 2025).

A plausible implication is that HVI-CIDNet+ is best understood as a second-stage extension of the original HVI/CIDNet paradigm: the HVI space provides the decoupled representation, the dual-branch CID backbone provides structured restoration, and the “plus” mechanisms inject priors specifically targeted at the failure regime where low-SNR local evidence is no longer sufficient.

4. Prior-guided and region-adaptive refinement

The Prior-guided Attention Block introduces DA-CLIP priors into branch interaction. For the I branch, a semantic prior is projected to a spatial tensor,

$\Delta = \mathbf{I}_{\max} - \min(\mathbf{I}_c)$3

and the query is computed as

$\Delta = \mathbf{I}_{\max} - \min(\mathbf{I}_c)$4

while keys and values are computed from HV-branch features,

$\Delta = \mathbf{I}_{\max} - \min(\mathbf{I}_c)$5

Cross-attention then updates the intensity feature map as

$\Delta = \mathbf{I}_{\max} - \min(\mathbf{I}_c)$6

followed by an MLP to produce the refined output feature (Yan et al., 9 Jul 2025). The semantic prior primarily promotes content restoration, whereas the degraded representation is injected into the HV side to emphasize degradation-aware color correction.

The Region Refinement Block is inserted in the I-branch decoder before CAB-based cross-branch fusion. It first predicts a partition mask that separates information-scarce from information-rich regions. Given an intermediate feature $\Delta = \mathbf{I}_{\max} - \min(\mathbf{I}_c)$7, an overview feature $\Delta = \mathbf{I}_{\max} - \min(\mathbf{I}_c)$8, and positional encoding $\Delta = \mathbf{I}_{\max} - \min(\mathbf{I}_c)$9, the module forms

$$h = \cos\left(\frac{\pi \mathbf{h}}{3}\right), \qquad v = \sin\left(\frac{\pi \mathbf{h}}{3}\right).$$0

predicts a mask $$h = \cos\left(\frac{\pi \mathbf{h}}{3}\right), \qquad v = \sin\left(\frac{\pi \mathbf{h}}{3}\right).$$1, and splits features into

$$h = \cos\left(\frac{\pi \mathbf{h}}{3}\right), \qquad v = \sin\left(\frac{\pi \mathbf{h}}{3}\right).$$2

The high-difficulty branch uses window-based self-attention for information-scarce regions, whereas the low-difficulty branch uses contextual multi-scale convolution for information-rich regions (Yan et al., 9 Jul 2025).

The convolutional path employs three parallel depthwise convolutions with kernel sizes 5, 7, and 9, followed by dynamic 3×3 filtering: $$h = \cos\left(\frac{\pi \mathbf{h}}{3}\right), \qquad v = \sin\left(\frac{\pi \mathbf{h}}{3}\right).$$3 This design assigns convolution to information-rich regions and self-attention to information-scarce regions, ensuring accurate brightness adjustments (Yan et al., 9 Jul 2025). In effect, the module treats extreme darkness as a region-wise heterogeneity problem rather than a globally uniform enhancement problem.

5. Optimization, datasets, and empirical performance

HVI-CIDNet+ is trained with a combined HVI-space and sRGB-space objective,

$$h = \cos\left(\frac{\pi \mathbf{h}}{3}\right), \qquad v = \sin\left(\frac{\pi \mathbf{h}}{3}\right).$$4

where the first term constrains the HVI distribution of the restored image and the second constrains pixel-level fidelity in sRGB (Yan et al., 9 Jul 2025). The original CIDNet literature further decomposes $$h = \cos\left(\frac{\pi \mathbf{h}}{3}\right), \qquad v = \sin\left(\frac{\pi \mathbf{h}}{3}\right).$$5 into $$h = \cos\left(\frac{\pi \mathbf{h}}{3}\right), \qquad v = \sin\left(\frac{\pi \mathbf{h}}{3}\right).$$6, edge, and perceptual components, and shows that combined HVI and RGB supervision is superior to either alone (Yan et al., 2024).

Training uses LOLv1, LOLv2-Real, LOLv2-Synthetic, SICE original, and a Sony-Total-Dark subset derived from SID. For LOL datasets, training uses $$h = \cos\left(\frac{\pi \mathbf{h}}{3}\right), \qquad v = \sin\left(\frac{\pi \mathbf{h}}{3}\right).$$7 crops with random rotations and flips; for SICE and SID, $$h = \cos\left(\frac{\pi \mathbf{h}}{3}\right), \qquad v = \sin\left(\frac{\pi \mathbf{h}}{3}\right).$$8 crops are used. Optimization uses Adam with $$h = \cos\left(\frac{\pi \mathbf{h}}{3}\right), \qquad v = \sin\left(\frac{\pi \mathbf{h}}{3}\right).$$9, $\mathbf{P}_\gamma$0, an initial learning rate of $\mathbf{P}_\gamma$1, and cosine annealing to $\mathbf{P}_\gamma$2. LOL training runs for 1000 epochs with batch size 4, while SICE and SID run for 500 epochs with batch size 8 (Yan et al., 9 Jul 2025).

On paired LLIE benchmarks, HVI-CIDNet+ reports stronger results than the original CIDNet. On LOLv1, CIDNet achieves $\mathbf{P}_\gamma$3 dB PSNR, $\mathbf{P}_\gamma$4 SSIM, and $\mathbf{P}_\gamma$5 LPIPS, while HVI-CIDNet+ reports $\mathbf{P}_\gamma$6 dB, $\mathbf{P}_\gamma$7, and $\mathbf{P}_\gamma$8, respectively. On LOLv2-Real, CIDNet reports $\mathbf{P}_\gamma$9, and HVI-CIDNet+ reports $T(\cdot)$0. On LOLv2-Synthetic, CIDNet reports $T(\cdot)$1, and HVI-CIDNet+ reports $T(\cdot)$2 (Yan et al., 9 Jul 2025).

The extension is substantially larger than the original model. The original CIDNet is reported at $T(\cdot)$3M parameters and $T(\cdot)$4G FLOPs, whereas HVI-CIDNet+ is reported at $T(\cdot)$5M parameters and $T(\cdot)$6G FLOPs (Yan et al., 9 Jul 2025). The gain is therefore not a lightweight refinement; it is a prior-heavy expansion targeted at the extreme-darkness regime.

Ablations isolate the contributions of the HVI representation, prior injection, and region refinement. Replacing HVI with sRGB or HSV degrades performance; in one reported color-space ablation on LOLv1, sRGB yields $T(\cdot)$7 dB PSNR, $T(\cdot)$8 SSIM, and $T(\cdot)$9 LPIPS, HSV yields $\gamma_G$0, and the full HVI-based model yields $\gamma_G$1. Adding degraded representations alone, latent semantic priors alone, both priors jointly, and then RRB progressively improves the CIDNet baseline, with the full model reaching the best reported values in that ablation (Yan et al., 9 Jul 2025).

6. Position within the decoupling literature

HVI-CIDNet+ belongs to a broader research trajectory on explicit color–intensity decoupling. The immediate predecessor is the original HVI/CIDNet line, which introduced HVI as a trainable low-light color space and CIDNet as a dual-branch Color and Intensity Decoupling Network with Lightweight Cross-Attention (Yan et al., 27 Feb 2025, Yan et al., 2024). A parallel line, BCNet, treats LLIE in CIELAB as a decoupled brightening-and-colorization problem, with separate lightness and chrominance decoders and explicit user-controllable chrominance guidance (Wang et al., 2023). This suggests that HVI-CIDNet+ is one member of a larger family of decoupling architectures, distinguished by its HVI geometry and extreme-darkness priors rather than by decoupling alone.

Subsequent HVI-based work expands different parts of the design space. VCR retains HVI as the front-end and adds Variance-aware Channel Filtering, Triplet Channel Enhancement, and Color Distribution Alignment to reduce channel-level inconsistency and misaligned color distributions (Cheng et al., 10 Mar 2026). ICLR keeps the HVI branch decomposition but replaces naive luminance–chrominance interaction with a Dual-stream Interaction Enhancement Module and a Covariance Correction Loss designed for weak inter-chrominance correlation and luminance-induced chrominance errors (Xu et al., 17 Nov 2025). RHVI-FDD introduces an RHVI transform and a Frequency-Domain Decoupling module with low-, mid-, and high-frequency experts, explicitly presenting itself as a plug-in improvement on top of HVI-CIDNet (Yang et al., 7 Apr 2026).

Other extensions push HVI decoupling into neighboring problems. CLE-RWKV reformulates LLIE as controllable enhancement with a continuous illumination variable $\gamma_G$2, uses HVI for noise-decoupled supervision, and supervises intensity and chromatic components from different targets (Han et al., 26 Mar 2026). pHVI-ISPNet applies robust HVI ideas to RAW-to-RGB night photography rendering, combining RAW-domain feature processing, wavelet-based feature propagation, dynamic loss coefficients, and feature-distribution matching losses for color constancy (Kınlı, 30 Apr 2026). In shadow removal, CFSR does not define HVI-CIDNet+, but it introduces a custom HVI color space, a learnable illumination nonlinearity, and a physics-constrained restoration architecture that also supports explicit color–intensity separation (Wang et al., 20 Apr 2026).

Taken together, these developments indicate that HVI-CIDNet+ is best viewed as a specific, prior-augmented stage in the evolution of HVI-based decoupling. The defining contribution is not merely the use of separate color and intensity branches, but the integration of an HVI representation with latent semantic priors, degraded representations, and region-adaptive refinement for severely information-scarce low-light regimes (Yan et al., 9 Jul 2025).