acceptodds
Under review as a conference paper at ICLR 2027

LoopLUT: 3D Lookup Tables with Progressive Region Refinement for Real-Time 4K Image Enhancement

Abstract

Color enhancement of 4K images must meet a quality target under a tight compute budget. Three-dimensional lookup tables (3D LUTs) dominate real-time enhancement because they decide at low resolution and apply a per-pixel lookup at full resolution. A single global LUT, however, is spatially invariant, so an underexposed shadow and a well-exposed region that share a pixel value receive identical corrections. Spatially heterogeneous demands cannot be expressed by such a mapping. We propose LoopLUT, a region-cascaded 3D LUT with progressive refinement. A global LUT performs the overall correction, followed by K-1 loop iterations. In each iteration a gating head predicts at low resolution the region that still needs correction, then builds a residual LUT from the color statistics of that region alone. The cascaded gates form a partition of unity, so the output is a per-pixel convex combination of the K lookup results. Fusion is therefore performed by the gates themselves, with no separate fusion module and no interpolation error accumulating across rounds. The decision stage runs at a fixed 256 × 256 resolution, independent of output resolution, so a 4K image costs only K pure lookups. Extensive experiments across four benchmarks show that LoopLUT improves PSNR by up to 2.81 dB over the strongest prior method, while keeping real-time throughput at 4K. The same decomposition also generalizes well to underwater enhancement datasets.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Color Enhancement, 3D LUT, Loop

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_v9UjOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces LoopLUT, a real-time image-enhancement method that overcomes the iteratively predicts regional gates and residual LUTs from a fixed-resolution representation, then combines multiple LUT outputs through per-pixel weights to provide spatially adaptive correction. The authors evaluate LoopLUT on LCDP, MIT-Adobe FiveK, Mobile-Spec, and the LSUI underwater enhancement benchmark. The method achieves strong quantitative performance, including PSNR improvements of 1 dB+ on FiveK and almost 3 dB on Mobile-Spec over the respective SOTA baselines.

Critical strengths
  • Interesting and well-motivated architecture design. The partition-of-unity formulation provides a principled way to combine region-specific LUTs without repeatedly applying LUTs to intermediate full-resolution images. Moreover, the loop for decision reuse previous steps and leads to a constant constrained burden.

  • Strong empirical results on both performance and efficiency. LoopLUT reports the best PSNR and SSIM on LCDP, the best values across all four reported metrics on MIT-Adobe FiveK and Mobile-Spec, and competitive results on LSUI. It process 4K images at 16.23 ms on the reported RTX 4050 Laptop GPU (Table 3). The authors also report Thoughput and device memory.

Critical weaknesses
  • More direct and intuitive ablation study. While the round-count ablation (Table 2) measures the combined effects of additional LUTs, spatial gates, and recurrent refinement. A controlled comparison against a non-recurrent multi-LUT architecture would be better to establish whether conditioning each round on the previous enhancement state provides a substantial advantage over one-shot spatial LUT prediction.
Questions for authors
  • I also wonder how the resolution size for decision affect the accuracy and latency, especially for different target processing resolutions.
4 comments
AuthorsOpenReview ↗

Response to Reviewer v9Uj: Feedback Ablation and Decision Resolution

Thank you for the encouraging assessment of our design, results, and efficiency. We agree that the two points you raised needed further experiments. We have added both to the revised manuscript, with the additions marked in purple. All table and figure references below refer to the revised version.

W1. A controlled comparison with a non-recurrent multi-LUT model

We agree that the round-count ablation alone does not isolate the contribution of feedback, since increasing the number of rounds also adds LUTs and spatial gates.

We therefore trained a non-recurrent three-LUT control. It retains the same backbone, basis LUTs, cascaded spatial gates, partition-of-unity structure, losses, training procedure, and 2.89 M parameters. The only change is that the gates and region statistics use the original low-resolution input instead of the previous round's low-resolution enhancement result. This keeps the LUT count and model capacity fixed while removing feedback from the updated image.

The results on the 218 LCDP test images at native resolution, averaged over seeds 42, 43, and 44, are:

ModelPSNR ↑SSIM ↑LPIPS ↓CIEDE2000 ↓
Single global LUT (K=1)22.650.80170.16877.72
Non-recurrent multi-LUT (K=3)22.420.80750.16697.99
LoopLUT (K=3)23.560.86560.15467.50

The average PSNR gain over the non-recurrent control is 1.13 dB, with gains of +1.14, +1.03, and +1.23 dB for the three seeds. These results support the contribution of feedback within this architecture: later rounds benefit from observing the updated enhancement result, with the number of LUTs and parameters held fixed.

We have added this comparison to Section 4.4, “Effectiveness of feedback,” and Table 3. Appendix D.4, Table 8, and Figure 12 provide the per-seed and per-image results.

Q1. How decision resolution affects accuracy and latency at different output resolutions

We retrained LoopLUT at decision resolutions r = 128, 192, and 384 and compared them with the main model at r = 256. We fixed the seed at 42 and used the same decision resolution for training and inference. Accuracy is evaluated on LCDP; total latency is reported for outputs at 1620 × 1080 and 4K.

rDecision GFLOPsPSNR ↑SSIM ↑Decision (ms)Total, 1620 × 1080 (ms)Total, 4K (ms)
1281.5923.080.84701.925.3715.61
1923.4823.190.85591.925.0316.00
2566.13523.480.86161.845.1016.23
38413.7123.860.87161.995.7017.55
51224.32——2.807.3119.34

Latencies were measured on the same GPU in fp16, using the median of 60 consecutive runs after warm-up, and normalized to the r = 256 values in Table 4. For r = 512, we report latency only: training with batch size 16 requires about 7.1 GB of GPU memory, exceeding the 6 GB available on our GPU.

Increasing r from 128 to 384 improves PSNR by 0.78 dB. In particular, moving from the default r = 256 to r = 384 adds 0.38 dB, while total latency rises from 5.10 to 5.70 ms at 1620 × 1080 and from 16.23 to 17.55 ms at 4K.

For a fixed r, the decision stage operates at the same resolution regardless of the output size; the full-resolution execution stage accounts for the increase in cost with output size. The results therefore show how decision resolution can be used to trade additional latency for better accuracy, while the decision-stage latency remains a small part of the total at 4K.

We have added these results to Section 4.4, “Effectiveness of the decision resolution,” and Appendix D.5, Table 9.

Thank you again for these suggestions. They helped us test the role of feedback more directly and clarify the accuracy–latency trade-off.

Dear Reviewer,

Thank you again for your valuable feedback and time. We have carefully addressed your comments and questions in our rebuttal.

We would greatly appreciate it if you could take a moment to review our responses and let us know whether any concerns remain. We would be more than happy to provide further clarifications or additional evidence during the discussion period.

Thank you very much for your time and consideration!

Thank you again for your time and constructive feedback.

As only a few hours remain in the author–reviewer discussion period, we would greatly appreciate it if you could spare some time to review our rebuttal and revised manuscript. We have addressed your concerns through additional analyses, experiments, and clarifications.

We would be happy to clarify any remaining questions and sincerely welcome any further feedback.

Thank you for your time and consideration.

Thanks to the authors for their rebuttal. With added additional experiments on the multiple non-recurrent LUT, my concerns are mostly addressed. Thus, I will maintain my positive position on this submission. I hope the authors could add this experiment in their future versions to enhance the completeness of the work.

Reviewer_b3o6OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes LoopLUT for high-resolution image enhancement. A global LUT is followed by regional residual LUTs, with each region selected using the previous low-resolution enhancement result. The gates also define the weights for combining the LUT outputs. All decisions run at 256 x 256, while the LUTs are applied to the original full-resolution image. Experiments cover three photographic enhancement datasets and one underwater dataset.

Critical strengths
  1. The design keeps the expensive decision stage at a fixed resolution while allowing different regions to receive different corrections. This is useful for mixed-exposure scenes. The reported 16.23 ms latency for a 4K image on an RTX 4050 Laptop GPU supports the practical motivation (Table 3).
  2. The spatial gating ablation provides direct evidence for local correction: replacing spatial gates with image-level scalar gates reduces PSNR by 3.08 dB on LCDP (Figure 6). Table 1 also reports clear improvements over the listed baselines on FiveK and Mobile-Spec. These results support the value of local adaptation, subject to clarification of the comparison setup.
Critical weaknesses
  1. The current ablations do not isolate the contribution of recurrent feedback. Figure 6 compares spatial and scalar gates, while Table 2 changes the number of rounds. Neither comparison keeps the spatial model fixed and removes only feedback from the updated enhancement result. This is an important gap because feedback is the main distinction from a spatially adaptive LUT mixture. A three-round model using a fixed initial state for later decisions would be a more direct control.
  2. The cross-dataset comparison is difficult to assess without the complete baseline protocol. Sections 4.1-4.2 do not state whether the baselines were retrained on each dataset or how their inputs and targets were prepared. This matters when interpreting the large differences on FiveK and Mobile-Spec. The main text refers to Appendix C for these details, but this appendix is absent from the PDF available for this review. The uncertainty concerns comparability; it does not establish that the reported results are incorrect. The reported LCDP improvement over CSEC is much smaller, at 0.13 dB (Table 1).
Questions for authors
  1. For the three-round model, is there a comparison that removes feedback from the updated enhancement result while retaining spatial gates and matching the backbone, LUT bases and losses? This would help separate the benefit of feedback from that of spatial gating.
  2. Were the Table 1 baselines retrained separately on each dataset? Please clarify the splits, FiveK target expert, preprocessing and checkpoint sources, or point to the relevant supplementary details. This would help assess whether the reported gains reflect comparable settings.
4 comments
AuthorsOpenReview ↗

Response to Reviewer b3o6: Feedback Ablation and Baseline Protocol

Thank you for the careful review and for recognizing the practical motivation and spatially adaptive design of LoopLUT. We agree that the original ablations did not isolate feedback and that the missing Appendix C made the baseline comparison difficult to assess. We apologize for the omission. The revised PDF includes the full appendix and a new feedback control, with additions marked in purple. All references below are to the revised version.

W1 / Q1. Isolating feedback while retaining spatial adaptivity

We ran the three-round control you suggested. It keeps the spatial gates and their cascade, backbone, basis LUTs, losses, training procedure, and 2.89 M parameters unchanged. Only the information used for later decisions changes: every gate and region statistic uses the original low-resolution input instead of the previous round's low-resolution enhancement result. Spatial adaptivity is therefore retained while feedback from the updated result is removed.

We evaluated all models on the 218 LCDP test images at native resolution, using seeds 42, 43, and 44. The table reports their three-seed means; the seed-42 LoopLUT model is the one in Table 1 (23.48 dB).

ModelPSNR ↑SSIM ↑LPIPS ↓CIEDE2000 ↓
Single global LUT (K=1)22.650.80170.16877.72
Fixed-input spatial control (K=3)22.420.80750.16697.99
LoopLUT (K=3)23.560.86560.15467.50

The average PSNR gain over the fixed-input control is 1.13 dB, with gains of +1.14, +1.03, and +1.23 dB across the three seeds. This supports the contribution of feedback when spatial gates, LUT count, and model capacity are held fixed. The control is called “One-shot multi-LUT” in Table 3.

The experiment appears in Section 4.4, “Effectiveness of feedback,” and Table 3, with per-seed and per-image results in Appendix D.4, Table 8, and Figure 12. Figure 6, Table 2, and Table 3 now separately examine spatial adaptivity, round count, and feedback.

W2 / Q2. Baseline training, preprocessing, and checkpoint sources

Not all baselines were retrained on every dataset. We now state three settings in Section 4.2 and mark each entry in Table 1: R, trained from scratch on the target training split; F, fine-tuned on that split from existing weights; and P, released pretrained weights used directly.

MethodLCDPFiveKMobile-Spec
HDRNetRPF
RUASPRF
LCDPNetPPF
AdaIntFPF
LightenDiffusionFPF
CSECPPF
DnLUTRRR
CanonCGTFPF
ShiftLUTFFF

P describes our use of a checkpoint, not its original training domain. For example, HDRNet and AdaInt use authors' FiveK-trained weights, whereas LCDPNet's FiveK evaluation uses its LCDP-trained checkpoint. These settings should be considered when interpreting the results.

Splits and preprocessing. Train/validation/test splits are LCDP 1,415/100/218, FiveK 4,176/250/497, and Mobile-Spec 140/20/40. FiveK uses ExpertC targets. LCDP uses the released image pairs. FiveK uses rendered RGB pairs with the short side set to 480 pixels and input/target sizes aligned by bicubic interpolation. Mobile-Spec uses only the RGB branch: 16-bit inputs are encoded as I8 = round[255 · ln(1 + 1000 · I16 / 65535) / ln(1001)], while targets retain their original 8-bit values. These details are in Appendix C.1. All methods are scored at native resolution using PSNR, SSIM, LPIPS (AlexNet), and CIEDE2000 (Appendix C.2).

Checkpoints and training. We trained from scratch when pretrained weights were unavailable (e.g., DnLUT) or fine-tuning performed poorly. Appendix C.4, Table 6 lists the official implementations and checkpoint sources, including HDRNet's expertC_cm2, AdaInt's AiLUT-FiveK-sRGB.pth, LCDPNet's trained_on_ours.ckpt, and CSEC's csec.ckpt. Appendix C.4 also documents method-specific training and model-selection settings.

Interpreting the margins. On FiveK, LoopLUT achieves 21.92 dB versus 20.77 dB for DnLUT, trained from scratch on the same split: a 1.15 dB gain. On Mobile-Spec, all nine baselines are trained or fine-tuned on the target training split; LoopLUT achieves 29.79 dB versus 26.98 dB for AdaInt, a 2.81 dB gain. These margins are relative to the target-trained runner-ups. On LCDP, we agree that the PSNR improvement over CSEC is smaller: 0.13 dB (23.48 versus 23.35 dB).

Thank you again for pointing out these gaps. The added control tests feedback directly, and the expanded protocol makes the conditions behind each comparison explicit.

Dear Reviewer,

Thank you again for your valuable feedback and time. We have carefully addressed your comments and questions in our rebuttal.

We would greatly appreciate it if you could take a moment to review our responses and let us know whether any concerns remain. We would be more than happy to provide further clarifications or additional evidence during the discussion period.

Thank you very much for your time and consideration!

Thank you again for your time and constructive feedback.

As only a few hours remain in the author–reviewer discussion period, we would greatly appreciate it if you could spare some time to review our rebuttal and revised manuscript. We have addressed your concerns through additional analyses, experiments, and clarifications.

We would be happy to clarify any remaining questions and sincerely welcome any further feedback.

Thank you for your time and consideration.

AuthorsOpenReview ↗

Gentle Follow-up on Our Rebuttal

Dear Reviewer,

Thank you again for your time and valuable feedback. We would like to kindly follow up on our previous response. We have carefully addressed your concerns through additional experiments, analyses, and clarifications, and we would greatly appreciate it if you could take a moment to review our rebuttal and updated manuscript. We would be happy to provide any further clarification if needed, and we sincerely welcome your additional feedback. Thank you very much for your time and consideration.

Best regards, The Authors

Reviewer_6bsCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes LoopLUT to address the issue that global 3D LUTs cannot differentially enhance identical colors at different spatial locations. The model first generates a global LUT, then iteratively predicts spatial gates based on the current low-resolution enhancement result, and uses statistics within gated regions to generate subsequent LUT residuals. Cascaded gates form non-negative, per-pixel, sum-to-one fusion weights; all LUTs are directly applied to the original image. The decision process is fixed at 256×256256\times256 resolution, while table lookups and weighted fusion are performed at full resolution. The authors validate the method on LCDP, MIT-Adobe FiveK, Mobile-Spec, and LSUI datasets, and provide experiments on spatial gating, number of iterations, and resolution scaling. The paper reports that the default three-round model processes 4K images in 16.23 ms on an RTX 4050 Laptop GPU, achieving approximately 62 FPS.

Critical strengths
  1. The design strikes a reasonable balance between spatial adaptability and high-resolution efficiency. The model shares a single low-resolution backbone computation, preserving the feedback process at low resolution while enabling local color adjustments through spatial weights. Each LUT directly queries the original input, avoiding repeated interpolation from serial transformations. The 4K speed reported in Table 3 supports its practical deployment value, representing one of the most compelling advantages of this work.

  2. The reported enhancement gains span diverse tasks and are not limited to PSNR improvements. In Table 1, LoopLUT achieves optimal results across four metrics on both FiveK and Mobile-Spec, outperforming the second-best methods by 1.15 dB and 2.81 dB in PSNR, respectively; LSUI results also provide initial evidence for underwater color correction. Given confirmed comparison protocols, these findings suggest that the architecture holds application potential beyond single-exposure correction scenarios.

Critical weaknesses
  1. The comparison protocols and baseline selection are insufficient to support the current claims of superiority. The main text does not adequately clarify whether each baseline was retrained on the corresponding dataset, what input processing and target images were used, or whether training and model selection procedures are comparable—information that primarily resides in Appendix C, which is currently inaccessible. Notably, DnLUT was originally designed for image denoising, while CanonCGT targets reference-image-driven color grading; thus, task adaptation and reference image selection must be clearly explained. Meanwhile, spatially aware 3D and 4D LUTs cited in the paper are more directly relevant to the local enhancement goal presented here, yet they were excluded from the main experiments. These gaps make it difficult to establish a clear advantage over the most relevant methods. Particularly on the primary benchmark LCDP, the PSNR lead over suboptimal methods is only 0.13 dB, and perceptual and color metrics are not optimal, making the reliability of comparisons especially critical.

  2. The existing experiments fail to isolate the core contribution of "progressive feedback refinement." Changing the gating size from (32×32)(32 \times 32) to (1×1)(1 \times1) mainly tests the usefulness of spatial variation; increasing the number of rounds simultaneously alters model capacity, lookup frequency, and feedback depth; and Figure 7 shows only a single example. These modifications do not sufficiently demonstrate that, under similar capacity and computational cost, predicting gating based on feedback from previous rounds is more effective than predicting multiple spatial weights at once or eliminating feedback altogether. The independent contributions of regional statistical pooling and auxiliary gating supervision have also not been disentangled.

  3. Furthermore, the explanation of "support set contraction" in Section 3.2 requires clarification: according to the definition (mk∈(0,1))(m_k\in(0,1)) provided in the paper, cumulative gating remains positive across all positions, meaning the non-zero support set typically still covers the entire image. While the multiplicative structure ensures pointwise weight reduction, this does not automatically imply that subsequent rounds focus exclusively on smaller error regions. This weakens the authors' argument that structural properties lead to "gradual focusing on remaining errors."

Questions for authors
  1. Clarification of assessment protocol: Please provide Appendix C to clearly define the division of each dataset, the input processing of FiveK and the target expert version, the training or fine-tuning methods of each baseline, and the reference image source of CanonCGT. Are the publicly available results cited in LSUI the same as the ones used in this paper in terms of division and evaluation criteria?

  2. Verification of feedback mechanism: When keeping the backbone, LUT quantity, spatial gating resolution, and training settings as much consistent as possible, removing the current enhancement state feedback or predicting all spatial weights at once, how does the performance change? Is there a comparison with the same protocol for spatial-aware 3D LUT and 4D LUT? These results will directly affect the judgment of the core contribution.

  3. Clarification of ablation control: Based on the average gain calculated in Figure 6, the three-round scalar gating model is approximately (23.48−3.08=20.40)(23.48 - 3.08 = 20.40) dB, which is lower than the 22.62 dB of the single-round model in Table 2. Are the two places using the same evaluation criteria? If so, why does it decrease significantly after adding two rounds? Especially, do (Lgate)(L_{\mathrm{gate}}) and the regional coverage constraint apply to the scalar gating control control with an inappropriate training objective?

  4. Clarification of region contraction definition: Does the "support set" in the text specifically refer to the non-zero support set or the effective region after threshold processing? If the latter, please provide the threshold and its definition, and explain whether there is evidence on the entire test set indicating that the gating indeed corresponds to the regions that have not been corrected in the previous round.

  5. Clarification of efficiency measurement: For 4K, 16.23 ms uses what numerical precision, batch size, and GPU synchronization timing method? Does it include preprocessing and data transmission? For the mobile end, 100–200 ms corresponds to what device and input resolution? These pieces of information will help determine the actual deployment scope.

  6. Can the average fusion weight represent the area of the region? Figure 1 infers the subsequent rounds of correction for the smaller regions based on the average weights of (0.65,0.20,0.15)(0.65, 0.20, 0.15). However, a spatially constant gating mechanism can also produce exactly the same average weights. On what basis do the authors interpret "decreasing average fusion weights" as "shrinking correction regions"? Is it necessary to adjust the interpretation of these statistics in Figure 1?

  7. How do the fusion paths in Figure 2 correspond to the formulas? In the figure, the global T1T_1 branch directly enters the summation, while the local MnM_n and TnT_n branches are multiplied; the main text uses partition weights AkA_k to mix the complete LUT output, or equivalently, uses cumulative gating CkC_k to weight the residuals DeltakDelta_k. Please clarify which calculation method is actually implemented and explain the missing partition weights or residual operations in the figure.

6 comments
AuthorsOpenReview ↗

Response to Reviewer 6bsC (1): Evaluation Protocol and Relevant Baselines

Thank you for the detailed review and for recognizing the efficiency and spatial adaptability of LoopLUT. Your comments identified gaps in both our evidence and its presentation. We apologize that Appendix C was missing from the review PDF. The revised PDF includes the full appendix, additional controls, and clearer definitions, with additions marked in purple. All references below refer to the revised version.

W1 / Q1. Comparison protocol and baseline adaptation

Not every baseline was retrained on every dataset. Section 4.2 and Table 1 now distinguish R (trained from scratch on the target training split), F (fine-tuned on that split), and P (released weights evaluated directly):

MethodLCDPFiveKMobile-Spec
HDRNetRPF
RUASPRF
LCDPNetPPF
AdaIntFPF
LightenDiffusionFPF
CSECPPF
DnLUTRRR
CanonCGTFPF
ShiftLUTFFF

P describes how we used the weights, not their original training domain. For example, HDRNet and AdaInt use FiveK-trained weights, while LCDPNet's FiveK evaluation uses its LCDP-trained checkpoint. Appendix C.4, Table 6 gives the official implementations, checkpoint sources, and method-specific training and model-selection settings.

DnLUT. We agree that its original task is denoising. We adapted the official pipeline to paired enhancement, using the degraded image as input and its enhancement target as supervision, with MSE loss, Adam, and batch size 8. Training lasts 100/60/40 epochs on LCDP/FiveK/Mobile-Spec; Appendix C.4 provides the learning rates, crop settings, and selection procedures.

CanonCGT. On FiveK we use the released E2E.pth; each image receives another image's ExpertC target as reference, assigned by a permutation without fixed points (seed 42). On LCDP and Mobile-Spec, we fine-tune the released model and retrieve the training-set target closest to the input using CLIP (ViT-B/32). No image uses its own target as reference. These reference-selection rules are now stated in Appendix C.4.

Splits, inputs, and targets. Train/validation/test splits are LCDP 1,415/100/218, FiveK 4,176/250/497, Mobile-Spec 140/20/40, and LSUI 3,423/428/428. LCDP uses the released pairs. FiveK uses rendered RGB pairs with ExpertC targets, short side 480 pixels, and bicubic alignment; pairs with more than 2% relative aspect-ratio mismatch are discarded. Mobile-Spec uses the RGB branch: inputs are encoded as I8 = round[255 · ln(1 + 1000 · I16 / 65535) / ln(1001)], with the original 8-bit targets retained. Appendix C.1 describes this preparation; C.2 specifies native-resolution evaluation without cropping or rescaling, using PSNR, SSIM, LPIPS (AlexNet), and CIEDE2000.

LSUI. Table 5 now clarifies that the methods were evaluated on the same 428 test images using the same PSNR and SSIM computation. UGAN and Ucolor were retrained; UIE-DAL and U-shape Transformer use released weights. These entries are evaluations under our stated protocol, rather than an assumption that published scores from different protocols are interchangeable.

W1 / Q2. Spatially aware LUT baselines

We added 4D LUT and LUTwithBGrid, a spatially aware 3D-LUT method, in Appendix D.7, Table 11. LUTwithBGrid represents the latter family; it is not the cited 2021 method, for which we did not have a public implementation. Both added methods use their official model code and losses, are trained on the same LCDP split for 56 and 53 epochs respectively, selected by validation PSNR, and evaluated on the same 218 test images.

MethodPSNR ↑SSIM ↑LPIPS ↓CIEDE2000 ↓
4D LUT20.460.78280.234211.70
LUTwithBGrid21.800.79070.17008.06
LoopLUT23.480.86160.15587.53

We agree that the LCDP PSNR lead over CSEC is modest: 0.13 dB. LoopLUT leads on PSNR and SSIM, while CanonCGT leads on LPIPS and CIEDE2000. On FiveK and Mobile-Spec, the 1.15 dB and 2.81 dB margins are over target-trained runner-ups: DnLUT trained on our FiveK split and AdaInt fine-tuned on Mobile-Spec. We now make these settings explicit so that the gains can be judged in context.

AuthorsOpenReview ↗

Response to Reviewer 6bsC (2): Feedback and Ablation Controls

W2 / Q2. Separating feedback from spatial gating and model capacity

We agree that the original gate-size and round-count ablations did not isolate feedback. We therefore trained the fixed-input three-round control you suggested.

The control retains the backbone, three LUTs, basis LUTs, 32 × 32 spatial gates and their cascade, losses, training procedure, and 2.89 M parameters. It also uses the same number of full-resolution lookups. Only the inputs to gate prediction and region statistics change: they use the original low-resolution input instead of the previous round's low-resolution enhancement result.

Results on the 218 LCDP test images at native resolution, averaged over seeds 42, 43, and 44, are:

ModelPSNR ↑SSIM ↑LPIPS ↓CIEDE2000 ↓
Single global LUT (K=1)22.650.80170.16877.72
Fixed-input control (K=3)22.420.80750.16697.99
LoopLUT (K=3)23.560.86560.15467.50

The average gain over the fixed-input control is 1.13 dB, with gains of +1.14, +1.03, and +1.23 dB across the three seeds. The seed-42 LoopLUT model is the Table 1 model (23.48 dB). Section 4.4 and Table 3 report this experiment; Appendix D.4, Table 8, and Figure 12 give per-seed and per-image results. The control is named “One-shot multi-LUT” in Table 3.

These results support the benefit of enhancement-state feedback within this architecture. The fixed-input control retains cascaded gates; it is not a separately designed parallel predictor of all spatial weights. The added spatially aware LUT comparisons are reported in Part 1 and Appendix D.7.

Regional statistical pooling and auxiliary supervision

The feedback control changes the information supplied to both gate prediction and regional pooling. It therefore measures their combined feedback contribution. Appendix B.3 explains how pooling supplies the residual generator with statistics of the selected region. We acknowledge that the current experiment does not independently quantify pooling or disentangle the two feedback paths.

For auxiliary gate supervision, we did train the same configuration with and without LgateL_{\mathrm{gate}} (Appendix D.6, Table 10). Their PSNR differs by less than 0.1 dB. With the auxiliary term, the gate's mean Spearman correlation with the remaining error increases from 0.135 to 0.372 in round 2 and from 0.054 to 0.267 in round 3. The round-3 area above cumulative-gate threshold 0.25 decreases from 41.5% to 8.1%. The evidence supports improved error alignment and tighter effective regions; positive alignment is still present without the term.

Q3. Why the scalar-gate control is below the single-round model

Your calculation is correct: the Figure 6 scalar-gate control is approximately 20.40 dB, below the 22.62 dB single-round model in Table 2. Both use the same 218 test images, native-resolution evaluation, and mean per-image PSNR. The difference is not an evaluation-protocol change.

The two models produce a global mapping in different ways. The single-round model predicts one LUT directly and has no gates or gate losses. The scalar control retains the three-round cascade and blends its LUTs using one gate value per image. It also retains LgateL_{\mathrm{gate}} and the coverage term, as in the spatial model.

For a spatial gate, coverage regularization acts on its spatial mean and permits correction to concentrate in selected regions. For a scalar gate, that mean is the scalar itself, so the same term regularizes the blending coefficient of the whole image. Keeping the objective unchanged therefore does not make its effect equivalent across the two gate forms.

We have clarified this in Section 4.4, “Effectiveness of spatial gating.” Figure 6 measures the combined effect of gate form and how the existing objective acts on it. Its 3.08 dB gap should not be read as a clean architecture-only effect or as the gain over a directly predicted global LUT. We have not established that the scalar control is optimal under an objective tailored to scalar gates.

The direct comparison with the single-round model is Table 2: 22.62 to 23.48 dB, a 0.86 dB gain. Table 3 separately reports a three-seed mean gain of 0.91 dB over a single global LUT and 1.13 dB over the fixed-input spatial control.

AuthorsOpenReview ↗

Response to Reviewer 6bsC (3): Region Definition, Fusion, and Timing

W3 / Q4. What contracts, and whether it follows the remaining error

You are right: because sigmoid gates are positive, the nonzero support generally remains the whole image. Our original use of “support contraction” was imprecise.

Section 3.2 and Appendix A.3 now define the thresholded effective region as Rk(ε)={x:Ck(x)>ε}R_k(\varepsilon)=\{x:C_k(x)>\varepsilon\}. Since Ck+1=Ckmk+1≤CkC_{k+1}=C_km_{k+1}\leq C_k, we have Rk+1(ε)⊆Rk(ε)R_{k+1}(\varepsilon)\subseteq R_k(\varepsilon) for any fixed ε>0\varepsilon>0. This guarantees nesting, not a strict decrease in area or exclusive selection of remaining-error regions.

We now assess error alignment empirically over all 218 LCDP test images, computing statistics per image at the 32 × 32 gate resolution and then averaging (Appendix D.6, Table 10). Here Ek−1E_{k-1} is the channel-averaged absolute error between the previous low-resolution enhancement result and ground truth.

StatisticRound 2Round 3
Mean Spearman correlation of gate and remaining error0.3720.267
Images with positive correlation80.7%75.2%
Effective area, threshold 0.195.3%42.1%
Effective area, threshold 0.2579.2%8.1%
Effective area, threshold 0.549.1%0.2%

At positions in the top 25% by gate value, remaining error is 1.36/1.32 times that at other positions in rounds 2/3; error reduction is 2.9/3.4 times as large. These statistics support a tendency to emphasize remaining error, rather than a structural guarantee or a property of every image. Figure 7 remains an illustrative example.

Q6. Mean fusion weights are not region areas

We agree that a constant weight map can have the same mean. The example values 0.65, 0.20, and 0.15 are mean partition weights AkA_k: average output shares from each round. They do not measure the area of Rk(ε)R_k(\varepsilon).

The revised Figure 1 caption makes this distinction and points to Table 10 for thresholded areas. Table 10 also reports within-image standard deviations of 0.240 and 0.082 for C2C_2 and C3C_3, respectively, showing spatial variation separately from their means.

Q7. Figure 2 and the implemented fusion

You are right that the original schematic omitted the conversion from gates to partition weights. This made the fusion path unclear.

The implementation follows Equation (2): every LUT queries the original full-resolution input, and every lookup output, including the global branch, is weighted by its partition weight before summation. For three rounds, A1=1−m2A_1=1-m_2, A2=m2(1−m3)A_2=m_2(1-m_3), and A3=m2m3A_3=m_2m_3. The weights are computed at low resolution and bilinearly upsampled; their unit sum is preserved.

The residual expression in Equation (3) is an algebraically equivalent identity, derived in Appendix A.4, rather than a separate execution path. We have clarified this correspondence in the Figure 2 caption. The direct global branch in the simplified drawing should not be interpreted as an unweighted full-LUT output.

Q5. Timing protocol and mobile deployment

The 16.23 ms 4K result uses an RTX 4050 Laptop GPU, fp16, batch size 1, with device synchronization after each forward pass. We report the median of repeated runs after warm-up under sustained load. The decision stage uses CUDA-graph replay; full-resolution execution uses 128-row tiles.

The timed interval includes downsampling to 256 × 256, the decision stage, LUT lookups, and weighting. It excludes image decoding/encoding and host–device transfer.

Appendix D.8, Table 12 adds same-session measurements:

4K precisionOn deviceWith uploadWith upload and download
fp1616.58 ms19.72 ms23.88 ms
fp3226.29 ms31.24 ms35.82 ms

The 16.58 ms value is a remeasurement, distinct from the original 16.23 ms. Upload includes conversion from an 8-bit host image; download includes 8-bit quantization and copying back. The 23.88 ms figure therefore includes both transfers, but still excludes image decoding/encoding.

The phone is a Redmi Note 12 Turbo, Snapdragon 7+ Gen 2, Android 13, running ONNX Runtime on CPU, four threads, fp32, without quantization or pruning (Appendix E). The illustrated input is a 1620 × 1080 LCDP image: 129 ms for decision, 52 ms for rendering, and 198 ms total, including thumbnail preparation. The reported 100–200 ms range varies with the photograph; it is not a mobile 4K timing claim.

Thank you again for the care taken in this review. The comments helped us distinguish what the structure guarantees, what the experiments support, and where the evidence remains limited.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.