- 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 and a dual-scale contrast map , 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 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 , but appendix Algorithm 2 updates each channel before the next pair reads it. Eq. 19 forces monotonicity on the residual tables at versus , while the prose calls an intensity-order constraint. , , , , and the low-pass 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 , , , , the 2D LUT size , 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 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 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 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
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.
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.