acceptodds
Under review as a conference paper at ICLR 2027

Beyond Color Remapping: Photometric–Detail Network for Low-Light Image Enhancement

Abstract

Lookup table (LUT)-based enhancement provides an explicit dark-to-bright mapping through indexing and interpolation. However, existing paradigms remain constrained by two key limitations. RGB-only indices cannot distinguish pixels with similar chromaticity but different illumination and contrast conditions, whereas channel-independent spatial lookup neglects cross-channel structural coherence. To address these limitations, we propose the Photometric-Detail Network (PD-Net), a two-stage framework that explicitly models photometric degradation and spatial detail separately. First, we develop a Dual-Condition Photometric LUT (DCP-LUT) that integrates RGB values, an illumination map, and a dual-scale contrast map into a 5D degradation-aware index. This index enables joint correction of brightness, color, and contrast. Second, we introduce a Channel-Pair Spatial Detail LUT (CSD-LUT) that improves texture consistency through channel-pair projection, rotation-cascaded dilated L-shaped lookup, and cyclic Gauss-Seidel propagation. Furthermore, we factorize each 5D LUT into a sum of rank-one tensors using canonical polyadic (CP) decomposition, reducing storage complexity from to without explicitly materializing dense grids. Evaluations on nine benchmarks under full-reference and no-reference protocols show that PD-Net achieves state-of-the-art performance while retaining a compact memory footprint and explicit lookup-based inference.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Low-light image enhancement, lookup table, photometric correction, spatial detail restoration, canonical polyadic decomposition

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_xiyzOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PD-Net, which is a two-stage lookup model for low-light enhancement. DCP-LUT builds a CP-factorized 5D table on RGB plus a U-Net illumination map DD and a dual-scale contrast map CC, read out by pentalinear interpolation. CSD-LUT adds stage-wise channel-pair 2D LUTs, four-orientation dilated L-shaped residual lookups, and cyclic Gauss-Seidel updates. The indexing claim is not positioned against existing 4D LUTs, the CP claim overreaches AttentionLUT [Fu et al., 2024], several references do not support the sentences that cite them, and the LOL state-of-the-art statement is not backed by the methods or the protocol that define this task.

Critical strengths

DCP-LUT gives the lattice coordinate, the boundary-safe index, and the 32-vertex weights. CSD-LUT gives the L-shaped offsets, the rotation update, and the cyclic write-back order. The condition network is a full-resolution U-Net with base width 8. Dropping both condition maps costs 0.82 dB PSNR on SICE, and replacing pairwise projection costs 0.53 dB. Table 4 and appendix Eq. 17 to Eq. 19 likewise separate pixel, photometric, smoothness, and monotonicity terms. Among LUT baselines on FiveK, Table 1 is internally consistent: PD-Net leads PSNR, SSIM, and ΔE\Delta E at both resolutions, uses 848.9K parameters against 11000K for FastLLVE and 10742K for WLUT, and lists a 3.2 ms runtime. That efficiency comparison is the strongest empirical result in the submission.

Critical weaknesses

Attribution is overstated where the method claims to be new. Same-RGB pixels with different context are already separated by 4D LUT [Liu et al., 4D LUT, 2022, attribution survey], which is absent from the bibliography. WLUT [He et al., 2026] is a 4D low-light LUT, but the text cites it as the SSIM loss (Sec. 2.5) and does not use it when Sec. 2.2 introduces condition indexing. Contribution 4 presents a CP-factorized representation as introduced here, while Eq. 11 cites that construction to AttentionLUT [Fu et al., 2024]. The low-light ranking does not support the abstract. Table 2 cites CIDNet as [Zhao et al., 2025], an underwater detector. The low-light CIDNet [Yan et al., HVI-CIDNet, 2025, attribution survey] reports 28.201 dB, 24.111 dB, and 25.705 dB PSNR on LOLv1, LOLv2-real, and LOLv2-syn, against 23.81 dB, 24.11 dB, and 25.13 dB in Table 2. Only LOLv2-real agrees closely, and GT-mean is never stated. Retinexformer [Cai et al., Retinexformer, 2023, attribution survey], at 27.140 dB on LOLv1 in the HVI-CIDNet table, is not compared. LPIPS is absent. Table 1 is FiveK retouching. Ablations never leave SICE. The formulas do not close. Eq. 10 writes Pt=Pt−1+RΦSt(Pt−1)P^{t}=P^{t-1}+R^{t}_{\Phi_{S}}(P^{t-1}), but appendix Algorithm 2 updates each channel before the next pair reads it. Eq. 19 forces monotonicity on the residual tables ΦS\Phi_{S} at λm=10\lambda_{m}=10 versus λs=10−4\lambda_{s}=10^{-4}, while the prose calls LmL_{m} an intensity-order constraint. NN, KPK_{P}, KSK_{S}, TT, and the low-pass BB in Eq. 17 are never set. Cited support is also mismatched: [Liu et al., 2023] is event-guided video, [Meng et al., 2026] is MRI, and [Liu et al., 2026] is not NIQE, BRISQUE, or PI.

Questions for authors

Please put NN, KPK_{P}, KSK_{S}, TT, the 2D LUT size NMN_{M}, the crop size, and the FiveK expert split in the main text. State the image resolution for the 3.2 ms entry in Table 1, and whether EθE_{\theta} runs at that resolution. A 4K full-resolution U-Net plus repeated 32-vertex lookups is not obviously a 3.2 ms operator on the reported GPU. Please reconcile Eq. 10 with Appendix Algorithm 2. Either redefine RΦStR^{t}_{\Phi_{S}} on the in-place state, or show that the written equality is only a symbol for the sum of the three updates. Also state why Eq. 19 should be monotone on residual spatial tables, and give the kernel of BB in Eq. 17. Please name the blind noise estimator and the sampling rule that produces n=444 in every violin plot of Fig. 4 to Fig. 6. Lower noise alone does not separate denoising from blur unless a sharpness or LPIPS number is reported beside it. These questions are clarification requests. Parallel propagation and video, named in the conclusion, are future work and do not affect the rating.

6 comments
AuthorsOpenReview ↗

Response on Experimental Protocols, Hyperparameter Sensitivity, and Evaluation: We clarify dataset splits, preprocessing, and training settings; justify LUT resolutions, CP ranks, and refinement stages; specify the inference timing protocol; establish consistency between Equation (10) and Gauss–Seidel updates; and detail the blind noise estimator, 444-patch sampling protocol, and added LPIPS results to distinguish noise suppression from excessive smoothing.

Q1.Thank you for the suggestion

1.Crop Size and Dataset Splits: RGB images are normalized to [0,1]. Training uses random 256×256 crops from FiveK 480p and original-resolution LOL images, with horizontal/vertical flips and 0°, 90°, 180°, 270° rotations. Adam uses β=(0.9,0.999), batch size 8, learning rate 10⁻⁴, and 10⁵ iterations. Testing uses full images. Outputs are clipped to [0,1], scaled by 255, rounded, and saved as 8-bit RGB PNGs. FiveK uses 4,500/500 training/test images (480p training; 480p/native 4K evaluation). LOLv1, LOLv2-Real, and LOLv2-Syn follow official splits of 485/15, 689/100, and 900/100. DICM (64), MEF (17), LIME (10), NPE (8), and VV (24) are no-reference test sets. SICE uses 360/15 for Section 3.3, "Ablation Study". FiveK uses Expert C as reference.

2.We added "Hyperparameter Sensitivity Analysis" in Section 3.3, "Ablation Study" (page 10; Figure 7), varying one parameter at a time. Figure 7(1): Increasing 5D LUT resolution N improves photometric mapping, but overly dense lattices weaken local constraints and generalization; thus N=33. Figures 7(3)–(4): Higher photometric/spatial CP ranks Kₚ/Kₛ improve correction and refinement, but excessive ranks increase redundancy and noise fitting; thus Kₚ=Kₛ=16. Figure 7(2): More residual stages T reduce structural errors, but excessive refinement repeatedly alters recovered details, causing error accumulation and saturation; thus T=3. Figure 7(5): Higher pairwise projection LUT resolution Nₘ improves cross-channel mapping, but parameters grow quadratically with diminishing gains; thus Nₘ=256. These settings balance accuracy, structural fidelity, and complexity; each pairwise LUT has a 256×256 lattice. The analysis and parameter rationale are given in Section 3.3, "Hyperparameter Sensitivity Analysis" (page 10), and Figure 7.

3.The 3.2 ms inference time was measured on FiveK 480p images with batch size 1.

4.We compared Equation (10) in main Section 2.3, "Channel-Pair Spatial Detail LUT (CSD-LUT)" (page 5), with Algorithm 2 in supplementary Section 1.1 (page 5). Both describe the same procedure; we clarified the aggregate residual operator ℛ and its relation to sequential Gauss–Seidel updates. At stage t, initialize Pcurᵗ=Pᵗ⁻¹. Three channel pairs sequentially perform pairwise projection and four-orientation L-shaped lookup, yielding residuals UR(t,4), UG(t,4), and UB(t,4). Each residual is immediately written back for subsequent branches. Here, ℛᵗ denotes the stage-t aggregate residual operator associated with Φₛ, while UR(t,4), UG(t,4), and UB(t,4) denote the three target-channel residuals after the fourth orientation. The aggregate residual is defined as ℛᵗ(Pᵗ⁻¹) = Cat(UR(t,4), UG(t,4), UB(t,4)) = Pcurᵗ − Pᵗ⁻¹. As each channel is updated once per stage, the write-backs yield Pᵗ = Pcurᵗ = Pᵗ⁻¹ + ℛᵗ(Pᵗ⁻¹), consistent with Equation (10). We clarified this correspondence before Equation (10) and in line 25 of Algorithm 2.

5.We have clarified the blind noise estimation method and sampling protocol used in Figures 4–6 in Section 3.2, "Experimental Results", under "Qualitative Comparisons" (pages 7–8). We also added LPIPS to Table 1 (page 6) and Table 2 (page 7) to assess perceptual quality and detail preservation. Specifically, we employ a wavelet-domain median absolute deviation (MAD) estimator for blind noise estimation. RGB images are first normalized to [0,1] and converted to luminance Y, with the estimated noise standard deviations scaled by 255. Overlapping 32×32 patches are then extracted at stride 16, including boundary-aligned windows. If the number of patches exceeds 12,000, uniformly spaced samples in raster order are retained to limit the count to 12,000. Patches are ranked by S=median(|ΔₓY|)+median(|ΔᵧY|), where ΔₓY and ΔᵧY denote horizontal and vertical luminance differences. Ranking is based on the ground truth for paired datasets and the input image for unpaired datasets. The 444 lowest-scoring patches are selected, with ties resolved in raster order. All methods use identical patch locations, and noise is estimated separately from their enhanced outputs to ensure consistent comparisons.To distinguish noise suppression from excessive smoothing, we further examine structural similarity and perceptual quality. The SSIM difference maps in Figure 5 reveal local structural errors, while the newly added LPIPS measures perceptual differences between enhanced images and ground truth. PD-Net achieves LPIPS scores of 0.064 and 0.054 on FiveK at 4K and 480p, respectively, and 0.033, 0.028, and 0.031 on LOLv1, LOLv2-Real, and LOLv2-Syn, outperforming the corresponding comparison methods. Together with the visual results in Figures 4–6, these findings show that PD-Net reduces residual noise while preserving sharp edges and fine textures, rather than merely lowering noise estimates through image smoothing. The sampling protocol, evaluation metrics, and results are detailed in the cited sections, figures, and tables.

AuthorsOpenReview ↗

Additional Clarification of Novelty for Further Evaluation

Further Clarification of PD-Net's Novelty and Technical Significance

Building on our responses above, we provide clarifications to distinguish PD-Net from existing LUT methods, highlight its novelty and technical contributions, and facilitate the reviewer's reassessment of the paper's research significance and value.

1. From context-aware lookup to degradation-aware photometric mapping.

Existing 4D LUTs use pixel-wise context to distinguish identical RGB inputs. Yet distinguishing context is not equivalent to identifying the photometric conditions needed for low-light correction. Enhancement depends on spatial exposure, local contrast, and regional intensity relationships; the same RGB value may therefore require different corrections.

DCP-LUT constructs a joint index from RGB, spatial illumination, and adaptively fused dual-scale contrast. Illumination describes exposure variation; fine- and coarse-scale contrast characterize local intensity changes and broader regional relationships. These complementary conditions guide correction according to photometric state rather than observed color alone. The contribution lies in a degradation-aware indexing mechanism that explicitly integrates complementary photometric factors, enabling adaptive correction under spatially varying illumination and contrast. Section 3.3 separately ablates illumination, contrast, and both contrast scales, supporting their roles.

2. From independent lookup to channel-correlated spatial restoration.

Pointwise LUTs cannot explicitly capture neighborhoods. Direct indexing of multichannel neighborhoods is expensive, while channel-independent spatial lookup neglects cross-channel structure. CSD-LUT addresses this trade-off through channel-pair projection, rotation-cascaded dilated L-shaped lookup, and cyclic Gauss--Seidel propagation.

Each pairwise projection converts complementary color channels into a scalar response; five spatial samples then index a 5D LUT. Cross-channel cues thus enter spatial lookup without increasing its dimensionality. Each rotated L-shaped lookup takes the preceding directional response as input, creating progressive spatial propagation rather than independent directional features. Sequential Gauss--Seidel updates write back each channel correction immediately, so later branches use the newest state within the same stage. Cycling the update order reduces fixed-order bias.

The innovation lies in this coordinated lookup operator: channel-pair conditioning, directional recursion, and within-stage cross-channel propagation jointly support structural restoration. Section 3.3 isolates these effects through single-channel, single-orientation, and synchronous-update controls.

3. Unified CP-factorized photometric and spatial mapping.

PD-Net also establishes a unified CP-factorized design for two high-dimensional LUTs with different indexing semantics and restoration goals. DCP-LUT maps photometric coordinates to RGB responses; CSD-LUT maps projected neighborhoods to scalar detail residuals. Both reconstruct the 32 vertices required for pentalinear interpolation directly from CP factors, avoiding dense 5D grids. Independent factors and ranks accommodate their different mapping capacities, while spatial factors are shared across orientations and residual stages.

The unified CP-factorized representation enables complementary photometric and structural mappings to operate efficiently within a compact, explicit LUT framework. DCP-LUT first establishes a corrected photometric state; CSD-LUT then restores spatial detail through directional and cross-channel interactions. The supplementary rank analysis examines capacity, restoration quality, and complexity.

4. Evidence and precise attribution.

The controlled experiments provide targeted evidence for PD-Net's methodological contributions. Ablations examine illumination, both contrast scales, channel-pair projection, rotational cascading, and sequential propagation. We have also added LPIPS and clarified blind noise estimation and sampling, providing structural and perceptual evidence beyond noise estimates.

Our responses further distinguish the proposed mapping design from existing context-aware 4D LUTs and low-rank LUT representations. We will align citations and performance statements with their corresponding evaluation protocols. These comparisons clarify the specific methodological distinctions underlying PD-Net's contributions.

In summary, PD-Net advances explicit LUT enhancement through three connected developments: degradation-aware photometric indexing, channel-correlated directional spatial propagation, and unified CP-factorized realization. Together, they extend LUTs beyond color remapping into a coherent process of photometric correction and structural restoration. We respectfully ask the reviewer to reconsider the paper's novelty and significance in light of our full rebuttal.

Thanks for the detailed response by the authors. This response has addressed most of my concerns. However, I found that some very relevant works should be surveyed and included in the revised paper, including:

[1]. Brateanu, Alexandru, et al. "Multinex: Lightweight low-light image enhancement via multi-prior retinex." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026. (Multi-prior Retinex LLIE extends beyond the cited LYT-Net and needs positioning.)

[2]. Yang, Xingxing, Jie Chen, and Zaifeng Yang. "Learning physics-informed color-aware transforms for low-light image enhancement." 2025 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 2025. (Physics-informed color-aware transforms match PD-Net’s photometric and color correction focus.)

[3]. Balmez, Raul, et al. "Isalux: Illumination and semantics-aware transformer employing mixture of experts for low light image enhancement." 2026 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE, 2026. (Illumination–semantics MoE Transformer is a recent non-LUT LLIE competitor to cite.)

If all these concerns could be addressed, I would like to raise my evaluation.

Reviewer_XkNxOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PD-Net, a two-stage LUT-based framework for low-light image enhancement. DCP-LUT builds the photometric mapping from RGB values, predicted illumination, and dual-scale contrast; CSD-LUT then refines details via pairwise channel projection, rotation-cascaded spatial lookup, and cyclic Gauss–Seidel updates. CP factorization keeps the LUT storage small. Across paired and unpaired benchmarks, the method matches or beats the baselines they compare against. Ablations show that the illumination and contrast conditioning complement each other, and that the spatial refinement components and auxiliary losses each help.

Critical strengths

1.By introducing a unified CP-factorized representation, the storage complexity is reduced from O(N^5) to O(KN). 2.The results they report hold up well on both paired and unpaired benchmarks, and the ablations back up the claim that the added pieces actually pull their weight.

Critical weaknesses

CSD-LUT uses projected features and residual-updated RGB values as lookup coordinates, but the paper does not specify the ranges of these coordinates or how out-of-range values are handled. The LUT resolutions, CP ranks, and number of refinement stages are also left unspecified, and these choices directly affect the model's behavior, size, and runtime.

Questions for authors

1.How does the CP rank affect image quality, model size, and runtime? Please compare several rank settings to show whether stronger compression causes a noticeable loss of detail or introduces visual artifacts. 2.CIDNet is used as a comparison baseline, but the paper itself is actually a work on underwater object detection — how should this be explained?

2 comments
AuthorsOpenReview ↗

Response on CP Rank Sensitivity and CIDNet Baseline Clarification: We added a sensitivity analysis of photometric and spatial LUT ranks, examining reconstruction quality, model size, and inference efficiency. The results show that rank 16 achieves a favorable balance between detail preservation and complexity, while insufficient or excessive ranks degrade performance. We also corrected the CIDNet citation to Yan et al. (CVPR 2025), without affecting the experimental results.

Q1. Thank you for the reviewer's suggestion. We have added a CP rank sensitivity analysis in Section 7, "Sensitivity Analysis of CP Rank" (page 12), of the supplementary material, along with Figure 5 (page 13), which presents the corresponding results. The details are as follows:

"To examine the effects of CP rank on image quality, model size, and inference efficiency, we vary the ranks of the photometric and spatial LUTs jointly, setting KΦ=KP=KS∈{4,8,12,16,20,24,28,32}K_{\Phi}=K_P=K_S\in\{4,8,12,16,20,24,28,32\} while keeping all other settings fixed. The results are shown in Figure 5. CP decomposition represents a 5D LUT as a sum of rank-one tensors, with the rank determining its expressive capacity. At low ranks, the limited number of components cannot adequately represent the nonlinear relationships among RGB values, illumination, and contrast, or distinguish different spatial neighborhood structures. Consequently, pixels with similar colors but different illumination conditions may receive similar corrections, while distinct texture patterns may produce similar spatial responses. This leads to insufficient local contrast, loss of weak textures, and incomplete edge recovery. Increasing the rank improves the representation of photometric variations and spatial structures, thereby reducing reconstruction errors. However, a higher rank also introduces more learnable parameters. Once the principal mapping relationships are sufficiently represented, further increases offer limited gains and may increase sensitivity to noise in the training data, affecting enhancement stability. Moreover, both the parameter count and the computational cost of CP vertex reconstruction grow linearly with rank. This cost is further amplified by the rotation-cascaded lookup and multi-stage refinement in the spatial LUT. Thus, increasing the rank does not necessarily improve enhancement quality but consistently increases computational overhead. The results indicate that KΦ=16K_{\Phi}=16 provides a favorable balance between photometric correction, spatial detail recovery, model size, and inference efficiency.“

Q2. The CIDNet baseline refers to Yan et al.'s HVI: A New Color Space for Low-Light Image Enhancement (CVPR 2025), rather than Zhao et al.'s CIDNet: Cross-Scale Interference Mining Detection Network for Underwater Object Detection (Knowledge-Based Systems, 2025). The shared name caused the citation error, which we have corrected without affecting the results or conclusions.

AuthorsOpenReview ↗

Additional Clarification of Novelty for Further Evaluation

Further Clarification of the Technical Novelty and Significance

Building on our responses above, we provide the following additional clarifications to further highlight the novelty and technical contributions of PD-Net, and to facilitate the reviewer's reassessment of the paper's significance and research value.

1. From color indexing to degradation-aware photometric mapping.

The fundamental limitation of conventional LUTs lies in the information used to determine their mappings. RGB values describe observed color, not the illumination and contrast conditions under which that color appears. Pixels with similar RGB values can therefore require different corrections, yet an RGB-only LUT assigns them essentially the same response.

Our Dual-Condition Photometric LUT (DCP-LUT) addresses this ambiguity by introducing spatial illumination and dual-scale contrast into the RGB coordinates. Illumination describes regional exposure variation; contrast captures fine local changes and broader intensity relationships. These cues play complementary roles: together they allow the lookup to respond to the photometric context rather than to color alone.

The key advance is an explicit degradation-conditioned mapping that jointly models illumination and dual-scale contrast, enabling adaptive correction under spatially varying photometric conditions.

2. From pointwise lookup to channel-correlated spatial restoration.

Pointwise LUTs cannot explicitly represent neighborhood structure. Spatial LUTs introduce local context, but direct indexing of multichannel neighborhoods leads to rapidly growing tables, whereas channel-wise processing discards useful structural relationships across RGB channels. The challenge is to capture both types of dependence with manageable lookup dimensionality.

Our Channel-Pair Spatial Detail LUT (CSD-LUT) addresses this trade-off through three coordinated operations. Channel-pair projection incorporates cross-channel cues before spatial lookup, avoiding a single table indexed by the full multichannel neighborhood. Rotation-cascaded dilated L-shaped lookup aggregates directional and longer-range information sequentially: each orientation operates on the response of the preceding one, rather than producing an independent feature for later fusion. Cyclic Gauss--Seidel updates then let subsequent channel operations access the latest refined states, while rotating the update order mitigates fixed-order bias.

The contribution is a coordinated spatial lookup mechanism that integrates directional aggregation, cross-channel dependencies, and sequential information propagation, extending explicit LUT operations beyond independent pixel corrections.

3. A coherent division between photometric correction and structural recovery.

The two stages address related but distinct aspects of low-light degradation. Brightness, color, and contrast require condition-dependent photometric correction; edges and textures require spatial and inter-channel interactions. Jointly encoding photometric conditions and spatial neighborhoods in a single LUT would substantially increase mapping dimensionality and computational cost.

PD-Net instead assigns each task an appropriate indexing mechanism. DCP-LUT first corrects the photometric state; CSD-LUT then refines spatial structure in that corrected state. The first provides a basis for restoration; the second overcomes the spatial limits of pointwise mapping. CP factorization provides compact representations for both stages, enabling efficient execution of high-dimensional lookup operations without materializing dense 5D tables.

In short, PD-Net advances explicit LUT enhancement along two connected axes: what information determines the mapping, and how spatial and cross-channel information propagates through lookup operations. Its novelty lies in a unified, problem-driven formulation that couples degradation-conditioned photometric mapping with coordinated spatial and cross-channel restoration.

The controlled ablations in Section 3.3 examine the complementary photometric conditions and the separate contributions of channel-pair projection, directional cascading, and sequential updates. Alongside our rebuttal analyses, they support the design's rationale and practical value.

We hope these clarifications provide a clear basis for assessing PD-Net's methodological novelty and research significance. We appreciate the reviewer's constructive comments and respectfully ask that PD-Net's technical contribution be reconsidered in light of the full response.

Reviewer_Rw3hOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PD-Net combines a context-conditioned photometric LUT with a spatial refinement stage for low-light enhancement. The photometric lookup uses RGB together with learned illumination and contrast maps. The refinement stage mixes channel pairs and applies spatial lookups across several orientations, with sequential channel updates. CP factorization keeps the 5D tables small. Results are reported on FiveK, LOL and several unpaired datasets, with component ablations on SICE and further SMID/SDSD results in the supplement.

Critical strengths
  1. A coherent treatment of two distinct limitations of pointwise LUT enhancement.** Context-conditioned indexing addresses ambiguity between equally colored pixels in different neighborhoods, while the second stage introduces spatial and channel interactions. This is a plausible and technically explicit architecture rather than merely increasing LUT dimensionality without adding contextual information (Sections 2.2–2.4).

  2. The reported gains are supported by several complementary evaluations and targeted ablations.** Tables 1–3 show favorable paired and unpaired results, and Table 5 removes individual conditioning and spatial-refinement components. For example, removing both contextual coordinates reduces SICE PSNR from 27.21 to 26.39 dB. If the experimental protocols are confirmed, the combination of image quality and compact parameter storage would be practically useful.

Critical weaknesses
  1. The comparison protocol is too incomplete to assess the reported advantage. Section 3.1 gives the optimizer and training length, but not the train/test partitions, the FiveK target expert, or the preprocessing and metric conventions. I could not determine which datasets and checkpoints were used for the unpaired tests, which baselines were retrained, or how methods from video enhancement, denoising and pan-sharpening were adapted. There is also a specific reference problem: Table 2 lists CIDNet, but Zhao et al. (2025) is an underwater object detection paper in the bibliography (p. 11, lines 586–588). A wrong citation would be straightforward to fix; the more important question is which implementation produced those results.

  2. The speed claim is similarly hard to evaluate. Table 1 gives 3.2 ms alongside accuracy results at both 4K and 480p, without saying which resolution was timed or whether the condition estimator and all refinement stages were included. The actual lattice sizes, CP ranks, number of stages and pair-projection resolution are also missing. These settings matter when 32 vertices are reconstructed for each lookup and the spatial branches are applied serially. Reduced LUT storage alone does not establish fast inference, and the full model has more parameters than several LUT baselines in the table. I need the complete configuration and a matched timing protocol to judge this tradeoff.

Questions for authors
  1. Please specify the splits, FiveK expert target, preprocessing, output clipping and metric code behind Tables 1–3, together with the training data/checkpoints for each baseline. Which CIDNet implementation was evaluated, and how were methods from other tasks adapted?

  2. What values were used for N, K_P, K_S, T and the pair-projection lattice size? At what resolution was 3.2 ms measured? Please clarify the timing procedure and whether it covers the complete network; a breakdown for the condition estimator and refinement would be useful.

  3. D and C are learned without direct illumination or contrast supervision. Is there evidence that their particular photometric interpretation matters, beyond providing two contextual features? A comparison with generic context coordinates using the same estimator and parameter budget would help establish this.

3 comments
AuthorsOpenReview ↗

Response on Experimental Reproducibility, Model Configuration, and Photometric Conditions: We clarify dataset splits, preprocessing, evaluation protocols, baseline configurations, and the CIDNet citation. We detail hyperparameter settings, inference timing, the condition estimator, and spatial refinement. Ablation results further demonstrate the complementary contributions of illumination and contrast conditions to enhancement performance.

Q1: Thank you for the suggestion. We have added these experimental details to "Section 3.1 Experimental Settings" (page 6) and supplemented the code archive with dataset, evaluation, and configurations.

1.Dataset splits. FiveK uses 4,500/500 training/test images, with 480p training and both 480p and native 4K evaluation. LOLv1, LOLv2-Real, and LOLv2-Syn follow official splits of 485/15, 689/100, and 900/100. DICM (64), MEF (17), LIME (10), NPE (8), and VV (24) are used for no-reference evaluation. SICE uses 360/15 for ablations in Section 3.3 "Ablation Study". FiveK uses Expert C as the reference target.

2.Preprocessing. Images are converted to RGB and normalized to [0,1]. Training uses random 256×256 crops from FiveK 480p and original-resolution LOL images, with random horizontal/vertical flips and 0°, 90°, 180°, and 270° rotations. We use Adam (β=(0.9,0.999)), batch size 8, learning rate 1×10⁻⁴, and 1×10⁵ iterations. Testing uses full images. Outputs are clipped to [0,1], scaled by 255, rounded, and saved as 8-bit RGB PNG files.

3.Evaluation metrics. The code archive provides pd_lut/eval_metrics.py. Paired metrics include PSNR (8-bit RGB), SSIM (11×11 Gaussian window, σ=1.5, RGB average), ΔE (CIEDE2000), and LPIPS (AlexNet). NIQE, BRISQUE, and PI use pyiqa. Dataset scores average per-image results. Evaluation supports checkpoints and saved outputs.

4.Baseline reproduction. We have added configs/baselines.json to the code archive, including newly added information on the training datasets, weight sources, checkpoint paths, and result directories for each baseline, along with configs/pdnet_checkpoints.json for the PD-Net weights. In Section 3.2 "Experimental Results", Table 1 uses FiveK with Expert C, Table 2 follows official LOL splits, and Table 3 evaluates baselines under their original training protocols. Evaluation and reproduction scripts are included.

5.The CIDNet baseline refers to Yan et al.'s "HVI: A New Color Space for Low-Light Image Enhancement (CVPR 2025)", rather than Zhao et al.'s "CIDNet: Cross-Scale Interference Mining Detection Network for Underwater Object Detection" (Knowledge-Based Systems, 2025). The shared name caused the citation error, which we have corrected without affecting the results or conclusions.

Q2:

1. We added "Hyperparameter Sensitivity Analysis" in Section 3.3 " Ablation Study" (page 10). And we have added Figure 7 (page 10), which shows single-factor sensitivity with the other parameters fixed. Balancing accuracy, structural fidelity, and complexity, we set N=33, Kₚ=Kₛ=16, T=3, and Nₘ=256.

2. The 3.2 ms inference time was measured on FiveK 480p images with batch size 1.

3. We have added a detailed description of the inference timing protocol to Section 4 ”Network Complexity and Inference Latency" (pages 6 and 8) of the supplementary material, together with the module-wise computational costs in Table 2 (page 6). Tests use an RTX 4060 Ti, fused LUT kernels, 256×256 inputs, batch size 1, and FP16. After 10 warm-ups, we average 50 GPU-synchronized runs; the median and 95th percentile differ by less than 3%. Timing covers condition estimation, DCP-LUT, all three CSD-LUT stages, and output clipping. Decoding, encoding, and CPU–GPU transfer costs are 0.98, 4.64, and 0.11 ms, respectively. Total latency is 3.2 ms: 0.12 ms (3.7%) for condition estimation, 0.16 ms (4.9%) for photometric mapping, and 2.92 ms (91.4%) for spatial refinement. Their sum indicates full-network coverage.

4. We have added a detailed description of the condition estimator to "Section 1 "Dual-Condition Photometric LUT (DCP-LUT)" (pages 1–2) of the supplementary material, together with the layer-wise network configurations in "Table 1 (page 2)". Its five-level U-shaped encoder–decoder generates coarse features Y₂ and fine features Y₀ through skip connections. From Y₀, 1×1 and 5×5 convolutions predict illumination and fine-scale contrast; a 9×9 convolution on Y₂ with fourfold upsampling predicts coarse-scale contrast. Pixel-wise gating fuses both contrast maps (supplementary Equations (1) and (2)). We have added detailed descriptions of channel-pair projection, four-orientation cascaded L-shaped lookup, and Gauss–Seidel updates for spatial refinement and cross-channel propagation to Supplementary Section 1.1 Channel-Pair Spatial Detail LUT (CSD-LUT) (pages 4–5). Algorithm 2 (page 5) provides the corresponding procedure.

Q3:

1. We have added a detailed explanation of this design to "Section 2.2 Dual-Condition Photometric LUT (DCP-LUT)", under "Illumination and Contrast Condition Modeling" (page 3). Illumination map D captures spatial exposure variations, while contrast map C combines local textures and regional structures, providing complementary photometric conditions. Section 3.3 "Ablation Study", under Network Architecture (page 9), reports the ablations in Table 5 (page 9).

AuthorsOpenReview ↗

Additional Clarification of Novelty for Further Evaluation

Further Clarification of the Technical Novelty and Significance

Building on our responses above, we provide the following additional clarifications to further highlight the novelty and technical contributions of PD-Net, and to facilitate the reviewer's reassessment of the paper's significance and research value.

1. From color indexing to degradation-aware photometric mapping.

The limitation of a conventional LUT is not simply its dimensionality, but the information available to its lookup. RGB values describe observed color, not the illumination and contrast conditions under which that color appears. Pixels with similar RGB values can therefore require different corrections, yet an RGB-only LUT assigns them essentially the same response.

Our Dual-Condition Photometric LUT (DCP-LUT) addresses this ambiguity by introducing spatial illumination and dual-scale contrast into the RGB coordinates. Illumination describes regional exposure variation; contrast captures fine local changes and broader intensity relationships. These cues play complementary roles: together they allow the lookup to respond to the photometric context rather than to color alone.

The key advance lies in selecting LUT indexing variables that explicitly characterize complementary aspects of low-light degradation, enabling photometric corrections conditioned on illumination and contrast. DCP-LUT jointly formulates illumination and dual-scale contrast as complementary photometric conditions, establishing an explicit mapping that adapts to spatially varying degradation.

2. From pointwise lookup to channel-correlated spatial restoration.

Pointwise LUTs cannot explicitly represent neighborhood structure. Spatial LUTs introduce local context, but direct indexing of multichannel neighborhoods leads to rapidly growing tables, whereas channel-wise processing discards useful structural relationships across RGB channels. The challenge is to capture both types of dependence with manageable lookup dimensionality.

Our Channel-Pair Spatial Detail LUT (CSD-LUT) addresses this trade-off through three coordinated operations. Channel-pair projection incorporates cross-channel cues before spatial lookup, avoiding a single table indexed by the full multichannel neighborhood. Rotation-cascaded dilated L-shaped lookup aggregates directional and longer-range information sequentially: each orientation operates on the response of the preceding one, rather than producing an independent feature for later fusion. Cyclic Gauss--Seidel updates then let subsequent channel operations access the latest refined states, while rotating the update order mitigates fixed-order bias.

The contribution lies in a coordinated lookup mechanism that integrates directional aggregation, cross-channel dependencies, and sequential information propagation for spatial detail restoration. It extends explicit LUT operations from independent pixel corrections to spatially coordinated detail restoration.

3. A coherent division between photometric correction and structural recovery.

The two stages address related but distinct aspects of low-light degradation. Brightness, color, and contrast require condition-dependent photometric correction; edges and textures require spatial and inter-channel interactions. A pointwise map cannot provide both, while placing all photometric conditions and neighborhood samples into one LUT would impose an impractical dimensional burden.

PD-Net instead assigns each task an appropriate indexing mechanism. DCP-LUT first corrects the photometric state; CSD-LUT then refines spatial structure in that corrected state. The first provides a basis for restoration; the second overcomes the spatial limits of pointwise mapping. CP factorization provides compact representations for both stages, enabling efficient execution of high-dimensional lookup operations without materializing dense 5D tables.

In short, PD-Net advances explicit LUT enhancement along two connected axes: what information determines the mapping, and how spatial and cross-channel information propagates through lookup operations. Its novelty lies in a unified, problem-driven formulation that couples degradation-conditioned photometric mapping with coordinated spatial and cross-channel restoration.

The controlled ablations in Section 3.3 examine the complementary photometric conditions and the separate contributions of channel-pair projection, directional cascading, and sequential updates. Alongside our rebuttal analyses, they support the design's rationale and practical value.

We hope this clarification makes the distinction from existing LUT approaches more apparent. We appreciate the reviewer's constructive comments and respectfully ask that PD-Net's technical contribution be reconsidered in light of the full response.

AuthorsOpenReview ↗

A Sincere and Earnest Request for the Reviewer's Response

We sincerely thank the reviewer for the careful evaluation and constructive suggestions. We have taken every concern seriously and made substantial revisions accordingly. In particular, we have added ablation experiments and hyperparameter sensitivity analyses, provided detailed experimental protocols and baseline configurations, clarified the complete inference timing procedure, and expanded the supplementary material with implementation details. We have also provided a further explanation of PD-Net's technical novelty and contributions.

We have made every effort to address all the reviewer's concerns and hope that these revisions provide sufficient clarification and experimental evidence. We sincerely hope the reviewer will consider our responses and reconsider the evaluation in light of these improvements. We deeply appreciate your valuable time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to Reviewers: Experimental Reproducibility, CIDNet Citation Correction, Hyperparameter and CP Rank Sensitivity, Inference Efficiency, Equation Consistency, Noise Evaluation, and Photometric Condition Validation etc.

Response to the Reviewers

We thank the reviewers for their suggestions. The revisions address experimental reproducibility, model settings, timing, residual updates, noise evaluation, and photometric conditions.

1.Experimental protocols and reproducibility. Section 3.1 (p. 6) and the archive document protocols. FiveK uses 4,500/500 images (Expert C; 480p train, 480p/4K test); LOLv1, LOLv2-Real, and LOLv2-Syn use official splits of 485/15, 689/100, and 900/100. No-reference tests: DICM (64), MEF (17), LIME (10), NPE (8), VV (24); SICE: 360/15 for ablations. RGB images are normalized to [0,1]. Training uses random 256×256 crops from FiveK 480p and original-resolution LOL images, horizontal/vertical flips, and 0°/90°/180°/270° rotations; Adam (β=(0.9,0.999)), batch 8, lr 10⁻⁴, 10⁵ iterations. Full-size outputs are clipped to [0,1], multiplied by 255, rounded, and saved as 8-bit RGB PNGs. Files include pd_lut/eval_metrics.py (PSNR: 8-bit RGB; SSIM: 11×11 Gaussian, σ=1.5, RGB average; ΔE: CIEDE2000; LPIPS: AlexNet; NIQE/BRISQUE/PI: pyiqa), configs/baselines.json (training sets, weight sources, checkpoints, result paths), and configs/pdnet_checkpoints.json. Scores average per-image results. Section 3.2, Tables 1–3, reports FiveK, LOL, and no-reference baselines under their original protocols.

2.CIDNet citation correction. The evaluated CIDNet is Yan et al.'s "HVI: A New Color Space for Low-Light Image Enhancement" (CVPR 2025), not Zhao et al.'s "CIDNet: Cross-Scale Interference Mining Detection Network for Underwater Object Detection" (Knowledge-Based Systems, 2025). We corrected the name-related citation; results remain unchanged.

3.Hyperparameters and CP rank. Main Section 3.3, "Hyperparameter Sensitivity Analysis" (p. 10; Fig. 7), varies one parameter at a time. N=33 balances mapping accuracy against sparse support in dense lattices (Fig. 7(1)); Kₚ=Kₛ=16 balances representation against noise fitting (Figs. 7(3)–(4)); T=3 avoids error accumulation from over-refinement (Fig. 7(2)); Nₘ=256 balances accuracy against quadratic growth, giving a 256×256 pairwise LUT (Fig. 7(5)). Supplementary Section 7, "Sensitivity Analysis of CP Rank" (p. 12; Fig. 5, p. 13), jointly tests Kᵩ=Kₚ=Kₛ∈{4,8,12,16,20,24,28,32}, measuring quality, size, and runtime. Low ranks underfit weak texture and local contrast; high ranks increase noise/artifacts and computation. Rank 16 provides the best reported balance.

4.Timing and architecture. Table 1 reports 3.2 ms for FiveK 480p at batch 1. Supplementary Section 4, "Network Complexity and Inference Latency" (pp. 6–8; Table 2, p. 6), uses an RTX 4060 Ti, FP16 fused LUT kernels, 256×256 input, 10 warm-ups, and averages 50 GPU-synchronized runs (median and 95th percentile differ by <3%). The forward pass includes estimation (0.12 ms), DCP-LUT (0.16 ms), three CSD-LUT stages (2.92 ms), plus clipping; decoding, encoding, and CPU–GPU transfers take 0.98, 4.64, and 0.11 ms separately. Supplementary Section 1 (pp. 1–2; Table 1) details the five-level U-shaped estimator: Y₀ predicts illumination (1×1) and fine contrast (5×5); Y₂ predicts coarse contrast (9×9, 4×4 upsampling), fused by pixel-wise gating (Eqs. 1–2). Section 1.1 (pp. 4–5; Algorithm 2) covers pair projection, four-direction lookup, and Gauss–Seidel updates.

5.Equation (10) and Algorithm 2. Main Section 2.3 (p. 5) and supplementary Section 1.1 (p. 5) agree. At stage t, Pcurᵗ=Pᵗ⁻¹; three pairs produce residuals after four orientations with immediate channel write-back. Denoting the aggregate residual operator by ℛᵗ (for Φₛ), ℛᵗ (Pᵗ⁻¹)=Cat(UR(t,4),UG(t,4),UB(t,4))=Pcurᵗ−Pᵗ⁻¹, and Pᵗ=Pcurᵗ=Pᵗ⁻¹+ ℛᵗ (Pᵗ⁻¹). Clarified before Eq. (10) and in Algorithm 2, line 25.

6.Noise sampling and perceptual fidelity. Main Section 3.2, "Qualitative Comparisons" (pp. 7–8; Figs. 4–6), details wavelet-domain MAD estimation. RGB is normalized to [0,1], converted to luminance Y, and estimated noise standard deviations multiplied by 255. Extract 32×32 patches at stride 16 including boundaries; if >12,000, uniformly subsample in raster order. Rank patches by S=median(|ΔₓY|)+median(|ΔᵧY|), using GT on paired data or the input on unpaired data. Keep 444 lowest (raster-order ties); methods share patch locations, with separate output noise estimation. SSIM difference maps (Fig. 5) assess structural errors. LPIPS added to Tables 1 (p. 6) and 2 (p. 7): 0.064/0.054 on FiveK 4K/480p and 0.033/0.028/0.031 on LOLv1/LOLv2-Real/LOLv2-Syn, supporting detail retention beyond smoothing.

7.Photometric interpretation. Main Section 2.2 (p. 3) defines D as spatial exposure variation and C as local and regional contrast. Main Section 3.3, "Network Architecture" (p. 9; Table 5), finds that removing D or C reduces PSNR by 0.43/0.50 dB; removing both reduces PSNR by 0.82 dB, SSIM by 0.014, and raises ΔE by 0.47. Single-scale contrast variants also underperform.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.