acceptodds
Under review as a conference paper at ICLR 2027

Shared but Not Substitutable: Frequency-Aware Color–Intensity Collaboration for Low-Light Image Enhancement

Abstract

We propose SPARNet, a frequency-aware framework for low-light image enhancement that explicitly models shared and component-specific information between chromatic and intensity representations in the HVI color space. Motivated by the observation that low-frequency HV and I components exhibit increasing similarity at coarser scales while remaining weakly predictable from one another under linear mappings, SPARNet learns shared and private representations to exploit cross-component dependencies without compromising their complementary characteristics. Specifically, a three-level wavelet pyramid captures multi-scale low-frequency context, while shared–private latent fusion enables adaptive information exchange and preserves component-specific representations through explicit regularization. The fused low-frequency context further guides directional high-frequency subband correction, facilitating illumination restoration, detail recovery, and noise suppression. Experiments on paired and unpaired low-light benchmarks show that SPARNet achieves the best results in eight of nine metric–dataset combinations across the LOL benchmarks and the lowest average NIQE on five unpaired datasets. Ablation studies and representation analyses further demonstrate the effectiveness of shared–private frequency collaboration.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Low-Light Image Enhancement, HVI Color Space, Shared-Private Representation Learning, Frequency-Aware Learning, Wavelet Transform, Color-Intensity Collaboration

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_a1XeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates supervised low-light image enhancement in the HVI color space, which decouples chromatic HV from intensity I. Addressing how color and intensity branches should coordinate during restoration, the authors propose SPARNet.


Controlled 3-level Haar DWT and Centered Kernel Alignment on 689 LOLv2-Real image pairs reveal that low-frequency HV and I components exhibit growing representation similarity at coarser scales. However, grouped ridge regression yields an out-of-fold R2≤0.0598R^2 \le 0.0598, demonstrating that low-frequency color and intensity share common spatial layout yet remain linearly non-substitutable.


A three-level wavelet backbone isolates low frequencies, where a shared–private cross-space low-frequency latent fusion module explicitly partitions features into shared and private subspaces via complementary gating.


The framework enforces three regularizers: shared alignment, intra-branch orthogonality, and private purity to prevent representation collapse and cross-branch leakage. Level-1 directional high-frequency subbands (HL, LH, HH) are selectively updated using fused low-frequency context and branch-specific features to suppress sensor noise while preserving sharp boundaries.

Critical strengths

The paper opens with a carefully controlled empirical study using spatial-shuffling and cross-image null controls alongside grouped cross-validation, establishing a quantitative foundation for the dual-pathway architecture.


The combination of soft complementary gating and three Frobenius-norm regularizers enforces a clean division of roles. Representation analyses using stage-wise CKA matrices verify that the separation margin rises from −0.157-0.157 to +0.506+0.506.


SPARNet achieves the top score in 8 of 9 metrics across the LOL family (LOLv1, LOLv2-Real, LOLv2-Synthetic), attains the lowest average NIQE (3.53) across five unpaired datasets, and reports evaluations on extreme sensor-noise data (Sony-Total-Dark).

Critical weaknesses

On the primary real-world benchmark LOLv2-Real, although SPARNet leads in PSNR (28.602 dB) and SSIM (0.893), its perceptual score (LPIPS 0.110) lags behind several existing methods, indicating a tendency toward slight over-smoothing.


On the Sony-Total-Dark dataset, SPARNet yields only a minor PSNR improvement of +0.126 dB over CIDNet. In severe low-photon regimes, level-1 high-frequency subbands remain corrupted by heavy signal-dependent noise that low-frequency guidance struggles to filter out.


The motivation rests on the failure of linear ridge regression (R2≤0.0598R^2 \le 0.0598). Because color and luminance exhibit inherently non-linear dependencies in complex scenes, evaluating only linear mappings leaves the degree of non-linear predictability unaddressed.

Questions for authors

On the LOLv2-Real benchmark, SPARNet achieves superior PSNR and SSIM, but its perceptual distance metric (LPIPS 0.110) lags behind existing methods such as GSAD (0.096) and Wave-Mamba (0.105). Does the strong low-frequency context guidance in Equation 8 cause slight over-smoothing in fine textures, and can this trade-off be improved by adjusting the perceptual loss weight wpercw_{perc} or relaxing the high-frequency gating?


On the extreme low-light Sony-Total-Dark dataset, the performance gain over CIDNet is relatively modest (+0.126 dB PSNR). In severe photon-starved regimes where level-1 high-frequency subbands are dominated by sensor noise, have the authors explored incorporating explicit wavelet shrinkage or a preliminary noise-filtering prior before gated subband correction?


The motivation analysis relies on linear ridge regression yielding an out-of-fold R2≤0.0598R^2 \le 0.0598 to claim that HV and I are non-substitutable. Did the authors evaluate non-linear predictability (e.g., using a shallow MLP, Random Forest, or Mutual Information)? Could the low R2R^2 reflect non-linear coupling rather than an intrinsic lack of cross-component predictability?


What are the inference latency, FLOPs, and parameter counts of SPARNet compared to CIDNet and Wave-Mamba on standard input resolutions? Does the three-level CSLF module with bidirectional transformer attention introduce substantial computational latency on edge or mobile devices?

1 comment

We thank the reviewer for the detailed analysis. We clarify the interpretation of the motivation study, connect the perceptual-quality concern to the existing ablation, and report the scope of available computational measurements.

1. Low-frequency guidance and the PSNR–LPIPS trade-off.

The fixed-architecture ablation in Table 5 shows that shared–private regularization improves LPIPS on LOLv2-Real from 0.1279 to 0.1100, alongside improved PSNR and SSIM. Thus, the constraints are compatible with improved perceptual similarity in the tested setting. The final LPIPS nevertheless remains worse than GSAD’s 0.096 and Wave-Mamba’s 0.105 in Table 1, which we recognize as a remaining limitation.

Eq. (8) retains the input high-frequency subband and adds a learned gated correction. It does not directly replace high-frequency information with the low-frequency context. The reported LPIPS gap alone therefore does not identify low-frequency guidance as its cause, although the learned update and reconstruction objectives may affect texture retention.

Appendix D specifies a perceptual-loss weight of 0.012 for LOLv2-Real. Changing this weight or the gating strength is a reasonable direction to investigate, but no systematic sensitivity study is reported. We will distinguish the observed perceptual gap from the unverified explanation of over-smoothing and avoid claiming that either adjustment has already improved the trade-off.

2. Extreme sensor noise and explicit denoising.

Sony-Total-Dark already tests this boundary: Table 2 reports 23.030 dB PSNR and 0.688 SSIM, corresponding to gains of 0.126 dB and 0.012 over CIDNet. These results show a measured improvement, while the small PSNR margin also exposes the limitation under severe noise.

The current model learns contextual, direction-specific corrections without an explicit wavelet-shrinkage stage or calibrated sensor-noise model. When fine signal is overwhelmed by noise, contextual guidance alone cannot guarantee its recovery. Shrinkage or a preliminary denoising prior is a relevant extension, but neither is evaluated in the submission. We will make this boundary explicit without presenting an untested extension as an established solution.

3. Nonlinear predictability and the scope of the motivation.

Our intended conclusion is limited linear substitutability under the tested fixed cross-image mappings. The controlled study provides two observations: paired CKA increases from 0.1550 at LL1 to 0.2633 at LL3 and exceeds the null controls, while image-grouped five-fold ridge regression yields out-of-fold R² no greater than 0.0598. Together, these observations motivate modeling shared structure alongside complementary representations.

We agree that low linear R² can coexist with nonlinear coupling. The main text and Appendix F.3 already specify the linear scope; the study does not evaluate MLP, Random Forest, or mutual-information alternatives. We will consistently qualify the “not substitutable” wording to avoid implying information-theoretic non-substitutability or a result about arbitrary nonlinear predictors. The usefulness of the shared–private modeling choice is additionally supported by the restoration ablation in Table 5.

4. Parameters, computation, latency, and deployment.

An existing profiling record for the three-stage implementation with level-1 subband correction provides the following absolute costs. The input is 400×600, batch size is 1, and execution uses FP32 on an NVIDIA RTX 4090, including DA-CLIP:

QuantityRecorded value
Trainable restoration and conditioning parameters22.63 M
Frozen DA-CLIP parameters in the loaded model246.21 M
Total loaded parameters268.84 M
Counted computation423.13 GFLOPs
Mean model-forward latency117.41 ms/image

The computation figure uses 1 MAC = 2 FLOPs for counted convolution, linear, and attention matrix-product operations; it is not a complete count of every operator. Timing averages 300 forwards and includes the prior encoder, while excluding disk I/O, host-to-device transfer, and external preprocessing. Frozen parameters are included in the loaded-model total.

These measurements provide the cost of the profiled implementation, rather than a standardized speed comparison against CIDNet and Wave-Mamba. Such a comparison requires the same resolution, hardware, precision, and timing boundary. No edge/mobile measurements are available in the submission, so the current evidence does not establish real-time deployment on those devices. We will report the profiling scope explicitly and keep deployment claims within the available evidence.

Reviewer_N5MVOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SPARNet for low-light image enhancement in HVI space. It combines a three-level wavelet pyramid, shared-private low-frequency fusion, and directional high-frequency correction. Controlled HV/I analysis motivates the design, with experiments on paired and unpaired benchmarks.

Critical strengths
  1. The motivation study includes spatial shuffling, image mismatching, and image-grouped regression, making the observed HV/I relationship more convincing.
  2. The model achieves competitive results across several benchmarks, especially LPIPS on LOLv1 and average NIQE on unpaired data.
  3. The regularization ablation links restoration gains with changes in feature alignment and separation.
Critical weaknesses
  1. Table 4 mainly compares CSLF with simple concatenation. This does not isolate the shared-private design from extra attention and model capacity.
  2. Eq. (7) retains a residual path outside the shared-private split. How much does restoration actually rely on each pathway?
  3. DA-CLIP provides pretrained content and degradation priors, but its contribution is not ablated.
  4. The related work discusses HVI-FourNet and HVI-FFNet, but neither appears in the comparisons
Questions for authors
  1. Does CSLF outperform cross-attention at matched capacity?
  2. What happens when either latent pathway or the DA-CLIP priors are removed?
  3. Are the small benchmark gains stable across training seeds?
1 comment

e thank the reviewer for the careful questions about capacity, pathways, and external priors. The fixed-architecture regularization control and an existing DA-CLIP ablation provide two complementary pieces of evidence.

1. Shared–private learning versus additional model capacity.

Table 5 isolates the effect of shared–private regularization with the architecture, parameter count, and auxiliary-prior setting unchanged. Enabling the constraints improves PSNR by 0.473 dB and reduces LPIPS by 0.0179 on LOLv2-Real. Therefore, this particular improvement cannot be attributed to additional parameters or a different prior encoder. In parallel, the shared–private leakage score decreases from 0.589 to 0.051, and the separation margin changes from −0.157 to 0.506.

These results support the usefulness of the representation constraints within CSLF. We agree that the concatenation comparison in Table 4 does not establish superiority over capacity-matched cross-attention; that specific comparison is not included in the submission. We will distinguish the demonstrated regularization benefit from this broader architectural question.

2. Shared/private pathways and the residual bypass in Eq. (7).

Eq. (7) implements residual refinement: the mixed branch representation supplies the base feature, and the shared latent together with the corresponding private feature determines the learned update. The shared path is intended to exchange common context, while the private paths retain component-specific information. The split therefore constrains the refinement pathway; the base feature also reaches reconstruction through the residual connection.

Importantly, Table 5 shows that the constraints measurably affect both restoration and representation organization despite this bypass. The residual connection does not make the constraints ineffective in the reported experiment. However, this control does not assign an independent contribution to each latent path, and the submission does not include pathway-removal measurements. We will describe the residual-refinement interpretation explicitly and keep the pathway claims within that scope.

3. Contribution of the DA-CLIP priors.

An existing controlled LOLv1 run removes the DA-CLIP extractor and conditioning projectors while keeping the backbone, seed, training settings, and evaluation procedure fixed. At each run’s highest-PSNR checkpoint, the no-prior model obtains 28.3426 dB PSNR and 0.0619 LPIPS, compared with 28.6677 dB and 0.0565 for the full model. The combined conditioning contribution is therefore +0.3251 dB PSNR and a 0.0054 LPIPS reduction in these runs.

This ablation establishes a benefit from the auxiliary conditioning in the tested setting. Separately, Table 5 establishes a regularization benefit while the prior setting is held fixed. These controls distinguish two contributions, although they do not establish how CSLF and capacity-matched cross-attention would compare without the prior. We will include the existing ablation and state this scope clearly.

4. HVI-FourNet and HVI-FFNet.

These are relevant HVI/frequency-domain methods. Our specific focus is the allocation and coordination of shared and component-specific factors in the low-frequency HV/I pathways, using complementary gating and explicit representation constraints, followed by directional high-frequency correction. We will sharpen the related-work discussion around these design choices.

Their absence from the quantitative comparison limits its coverage. We will qualify ranking statements as applying to the tabulated methods; the current results do not support an empirical superiority claim over either omitted method.

5. Stability across training seeds.

The benchmark tables report single-run point estimates, as stated in Appendix E.3 and Appendix G. For interpreting the method, the fixed-architecture ablation (+0.473 dB PSNR with accompanying representation changes) provides a more direct control than an isolated small benchmark lead. Nevertheless, neither that ablation nor agreement across metrics replaces independent training repetitions. Seed stability and statistical significance are not established by the submitted results, and we will avoid treating small ranking margins as decisive evidence.

Reviewer_bxivOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents SPARNet, a HVI-based frequency-aware low-light image enhancement (LLIE) framework. Motivated by the empirical observation that low-frequency chromatic (HV) and intensity (I) components are correlated but not linearly predictable across scales, SPARNet proposes a structured approach to cross-component information exchange in HVI space, and extensive experiments on paired and unpaired benchmarks demonstrate promising performance across most metrics.

Critical strengths

1.The exploration of color–intensity collaboration in HVI colour space is interesting and well-motivated.

2.Experimental results on various paired and unpaired LLIE datasets show consistent perform gain, and ablation studies systematically verify the contribution of each component.

Critical weaknesses

1.In terms of numerical evaluation, the reported metrics on benchmarks (e.g., Tab. 1.) seems largely different than those in their original papers. For example, the CIDNet on Lolv2-real is 24.111 PSNR and 28.139 in the submission. It is not clear what are the specific reasons and more clarifications are needed.

2.It is also encouraged to augment more analysis on the capability boundary of SPARNet on extreme challenging cases such as strong noise or very dark/low-light cases.

3.The ablation studies are conducted on LOLV1 where the metrics are quite saturated and the gain of each module is somewhat marginal. It is suggested to perform ablation on more representative data to highlight the differences.

4.The benefit of introducing a frozen DA-CLIP encoder as auxiliary prior could also be further analyzed.

Questions for authors

Please see the weakness section above.

1.The reported metrics on benchmarks (e.g., Tab. 1.) seems largely different than those in their original papers. For example, the CIDNet on Lolv2-real is 24.111 PSNR and 28.139 in the submission. It is not clear what are the specific reasons and more clarifications are needed. 2.More analysis on the capability boundary of SPARNet on extreme challenging cases such as strong noise or very dark/low-light cases. 3.The ablation studies are conducted on LOLV1 where the metrics are quite saturated and the gain of each module is somewhat marginal. It is suggested to perform ablation on more representative data to highlight the differences. 4.The benefit of introducing a frozen DA-CLIP encoder as auxiliary prior could also be further analyzed.

1 comment

We thank the reviewer for recognizing the motivation for color–intensity collaboration and the consistent improvements across benchmarks. We clarify the evaluation setting and provide an additional ablation from existing experiment records below.

1. Source of the CIDNet score and evaluation protocol.

The CIDNet value of 28.139 dB in Table 1 matches the GT-mean-adjusted result of 28.1387 dB in its official repository. The 24.111 dB cited by the reviewer matches the separately listed result of 24.1106 dB without GT-mean adjustment. The source is the official CIDNet result table.

Our LOL evaluation implementation also applies GT-mean brightness adjustment: the prediction is rescaled using the ratio of reference and predicted mean grayscale brightness before metric computation. This explains the evaluation-setting distinction relevant to the CIDNet example. We should have labeled the setting explicitly. We will identify GT-mean-adjusted entries and document each baseline’s result source and protocol. Reference-assisted brightness adjustment must be distinguished from direct-output evaluation when interpreting the comparisons.

2. Performance boundaries under extreme darkness and strong noise.

The paper already evaluates this regime on Sony-Total-Dark (Table 2): SPARNet achieves 23.030 dB PSNR and 0.688 SSIM, improving over CIDNet by 0.126 dB and 0.012, respectively. Thus, the evaluation includes severe low-light conditions, although the PSNR gain is modest. The model uses low-frequency context to guide directional high-frequency updates, but lacks an explicit calibrated sensor-noise model. Recovery remains difficult when high-frequency signal is overwhelmed by noise. We will make this existing limitation more prominent and distinguish improvements in the reported benchmark from a general claim of robustness to arbitrary extreme noise.

3. Evidence beyond LOLv1.

Table 5 provides a controlled ablation on LOLv2-Real with the CSLF architecture and parameter count unchanged:

VariantPSNR ↑SSIM ↑LPIPS ↓
Without SP regularization28.1290.88480.1279
Full model28.6020.89300.1100

The improvements of 0.473 dB PSNR, 0.0082 SSIM, and 0.0179 LPIPS reduction demonstrate the utility of the constraints on the real-world benchmark without adding capacity. Table 4 additionally reports a 0.805 dB PSNR improvement for the full model over the concatenation baseline on LOLv1. We will distinguish these complementary controls more clearly: Table 4 studies the component combination, while Table 5 isolates regularization on LOLv2-Real. A full component decomposition on that benchmark remains outside the submitted evidence.

4. Contribution of DA-CLIP.

An existing LOLv1 ablation directly addresses this concern. The backbone, training settings, random seed, and evaluation procedure are the same; the ablated model removes the frozen DA-CLIP extractor and its conditioning projectors. All reported metrics below come from each run’s highest-PSNR checkpoint:

VariantPSNR ↑LPIPS ↓
Without DA-CLIP and its projectors28.34260.0619
Full SPARNet28.66770.0565

Adding this auxiliary conditioning improves PSNR by 0.3251 dB and reduces LPIPS by 0.0054. The model retains competitive restoration quality without the encoder, while the prior provides an additional benefit. This ablation measures the combined effect of the extractor and projectors; it does not separate content and degradation conditioning. We will include this result and clarify DA-CLIP’s auxiliary role.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.