{"id":"b89d326d-f820-4c3b-9287-c5b71cfe2cff","arxiv_id":"2607.22801","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A compact depth-conditioned YCbCr 4D lookup-table network enhances underwater images at 7–11 ms per frame while matching or exceeding much larger baselines in restoration quality.","lead":"This paper describes DY-LUT, a real-time underwater image enhancement system that stores corrections in depth-aware YCbCr lookup tables and learned per-pixel indices, claiming 9–304x speedups over larger deep models with comparable quality. It matters for deploying enhancement on robots, sonar systems, and edge devices where latency and power are limited.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-set driven design selection on UIEB-90 may inflate the reported quality: hyperparameters, LUT capacity, and ablations are all chosen on the same test set used for final scoring, with no validation split or error bars.","rationale":"The reader's formal weakest assumption is depth reliability, but the rationale also identifies the model-selection-on-test-set risk. I consider the selection-on-test issue more load-bearing for the central claim, because it directly affects the 'competitive quality' headline and the relative ranking against baselines; depth robustness is actually supported by Tab. 3 across several sources. The concern is not fatal: the architecture and speedups are plausible, and the reported quality may survive a proper validation protocol. It does, however, strengthen the conditional: the paper should either provide a validation-split re-run or explicitly disclose that the reported scores are after selection on UIEB-90. Since the reader already gave CONDITIONAL, I keep that verdict unchanged.","tokens_in":15395,"tokens_out":4582,"duration_ms":55017,"concrete_test":"Use an 80/20 split of UIEB-800 (or a separate validation set from UIEB-890 excluding UIEB-90) and re-run the design grid: K in {1..5}, N in {17,25,33}, the Tab. 8 module/loss ablations, and the Tab. 7 coordinate variants. Select the best configuration on validation only, then train it once on the full UIEB-800 and evaluate on UIEB-90. If the resulting PSNR drops by more than ~0.5 dB from 23.245 or the ranking versus SMDR-IS/HCLR-Net changes, the headline quality claim is inflated by test-set selection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that DY-LUT is quality-competitive with 3.56M parameters and a large speedup. The evidence for the quality half rests on a selection-on-test problem. UIEB-90 is used both as the full-reference test set and as the criterion for design choices: Supplement D selects K=3 and N=25 by comparing PSNR on UIEB-90 (Tables 9/10); Tab. 8 chooses architecture and loss components by UIEB-90 PSNR; Tab. 7 selects the coordinate representation; the loss weights in Eq. 6 are likewise fixed and ablated against UIEB-90. No held-out validation split or multiple-seed variance is reported. Every headline number is therefore a best case after multiple looks at the test set, while the published baselines are evaluated with their own fixed configurations. This can inflate both the absolute 23.245 dB and the relative advantage over HCLR-Net/WF-Diff, which is the load-bearing part of the claim. The paper's Sec. 5 limitation discusses depth sensitivity but does not disclose or mitigate this selection effect, and the reader's own rationale flags the same risk.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DY-LUT, a 3.56M-parameter enhancement network that converts the input to YCbCr, combines it with an external depth map and a Sobel gradient, predicts image-level LUT fusion weights and a two-channel per-pixel degradation-index field, and uses these to query three learnable 4D LUTs with quadrilinear interpolation, followed by local refinement. It is trained on UIEB-800 and evaluated on UIEB-90 and LSUI for full-reference quality, U45 and UIEB-C60 for zero-shot no-reference quality, UIQAD for 1080P/4K latency, RUOD for detection, and FLSea for SIFT matching. The paper reports 23.245 dB PSNR on UIEB-90, best PSNR/SSIM on zero-shot LSUI, 9–304× speedups over high-capacity baselines, and ~7 ms adaptive 4K inference.","tokens_in":15749,"tokens_out":6854,"duration_ms":74178,"significance":"If the reported numbers hold in a properly controlled evaluation, the paper would make a useful contribution: it demonstrates that a compact, depth-conditioned LUT formulation can approach heavyweight restoration models at interactive rates, and the evaluation goes beyond a single benchmark (zero-shot, high-resolution, depth-source robustness, downstream tasks). The supplement's exact quadrilinear interpolation, pseudocode, and capacity sweeps are a strength. The central evidence, however, has two load-bearing gaps: model selection is performed on the same test set used for the headline numbers, and the main comparison gives DY-LUT a depth modality not available to baselines. These must be addressed before the quality-competitiveness claim can be accepted.","major_comments":[{"comment":"Design selection is circular with the headline evaluation. UIEB-90 is used as the full-reference test set, and all design choices are validated on UIEB-90: K and N are chosen from PSNR on UIEB-90 (Supplement Tables 9 and 10); architecture and loss components are ablated on UIEB-90 (Table 8); the coordinate representation, including the depth and CDI axes, is selected on UIEB-90 (Table 7); and the loss coefficients in Eq. (6) are fixed without a held-out set. Moreover, no multiple-seed variance is reported. Thus the 23.245 dB PSNR, and the margins over HCLR-Net and WF-Diff, are best-case numbers obtained after repeated looks at the test set. This is load-bearing for the central claim of 'competitive quality.' Please introduce a proper validation split (or cross-validation), report mean±std over seeds, and either re-report or clearly qualify the headline results.","section":"Sec. 4.6; Supplement D; Eq. (6)"},{"comment":"The main comparison is not input-equivalent. DY-LUT receives an additional depth channel, X = cat[Y,Cb,Cr,D,Γ], while all baselines in Table 1 are RGB-only methods. Table 7 shows that adding depth is worth 1.69 dB in RGB and 0.86 dB in YCbCr (fixed chroma coordinates), and the final learned-coordinate design reaches 23.245 dB. Therefore part of the reported advantage over RGB-only baselines may be attributable to the extra modality rather than to the proposed LUT mechanism. Please include a no-depth DY-LUT row in the main tables (e.g., the YCbCr row of Table 7, 20.87 dB on UIEB-90) or provide depth-conditioned baselines, and phrase the quality comparison in terms of a fair protocol. The speed comparison is unaffected, but the quality claim needs this controlled comparison.","section":"Sec. 4.2, Table 1; Sec. 3.2; Table 7"}],"minor_comments":[{"comment":"The text says the averages are reported in 'Tab. 5 of the main paper,' but the downstream results are in Table 6 of the main paper.","section":"Supplement F.2"},{"comment":"The caption of Figure 1 says 'RGB-only end-to-end cost is reported in Sec. 4.3,' but Section 4.3 reports latency with Lite-Mono rather than an RGB-only quality number. Clarify early on that the main quality tables use externally supplied depth and that no-depth quality is available only in Table 7.","section":"Sec. 4.3, Fig. 1 caption"},{"comment":"The gradient loss is written as ||∇Y − ∇Ygt||1; the first Y should be Ŷ to match the reconstruction-loss notation. Minor notation issue.","section":"Algorithm 2, line 14"},{"comment":"For reproducibility, state explicitly whether code and trained checkpoints will be released. The supplement says 'all checkpoints, datasets, and evaluation settings follow Sec. 4.1,' but no release link is given.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the selection-on-test-set circularity plus the privileged depth input. Both are fixable: a held-out validation split (with error bars) and a no-depth DY-LUT comparison in the main tables would let the quality claim be assessed fairly. If after that the core numbers degrade substantially, the paper should be reframed as an efficiency-oriented contribution rather than a quality-competitive one."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"DY-LUT is a well-executed engineering contribution: a 3.56M-parameter network that combines a YCbCr 4D lookup table, scene depth as a fourth axis, and two jointly learned latent degradation indices. The ablations make a convincing case that each piece earns its place — YCbCr beats RGB, depth helps, and the learned coordinates are worth ~1.5 dB over fixed chrominance. The efficiency numbers are plausible for a LUT-based method, and the evaluation is unusually broad: zero-shot generalization, depth-source robustness, high-res latency on several devices, and downstream detection/matching.\n\nThe main soft spot is one the stress-test flag correctly: UIEB-90 is used for both design selection and headline scoring. LUT capacity (K,N), loss weights, and every architectural choice are tuned on UIEB-90 PSNR, then reported as the final number, with no validation split or error bars. That makes 23.245 dB a best case after multiple looks at the test set, and the margin over HCLR-Net and WF-Diff is correspondingly suspect. The zero-shot LSUI results mitigate this somewhat, since that dataset was not used for ablation, but they don't fix the selection problem on UIEB-90 itself.\n\nSecond, the main comparison is not apples-to-apples: DY-LUT receives an external depth map while the baselines do not. The paper is transparent about this, and Table 7 quantifies depth's contribution, but the headline tables would be fairer with an RGB-only DY-LUT row alongside the depth-conditioned version. Third, no code or data is released, which hurts reproducibility even though the method is described in enough detail to reimplement.\n\nNone of this invalidates the core claim. The depth-source robustness study is a good piece of work, and the paper is honest about its limitations — depth reliability in turbid water, adaptive downsampling trade-offs, and the non-identifiability of the latent indices. It deserves a serious referee: the idea is novel enough and the evaluation broad enough that the field would benefit from seeing it in print, but the authors should be pushed to report held-out validation or at least error bars, add an RGB-only baseline to the main comparison, and release code.","headline":"Solid engineering contribution with a real test-set selection problem and an unfair baseline comparison; the core LUT design is novel and worth refereeing.","tokens_in":16243,"tokens_out":2764,"would_cite":true,"duration_ms":29194,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Underwater image enhancement can be reduced to a depth-conditioned lookup in YCbCr space: a 3.56M-parameter network built from 4D lookup tables matches the restoration quality of far heavier models while running 9–304× faster and processing","keywords":["underwater image enhancement","lookup tables","4D LUT","YCbCr color space","depth-conditioned restoration","Beer-Lambert attenuation","real-time inference","adaptive downsampling"],"falsifier":"A concrete test: take a benchmark of turbid or textureless underwater images, run DY-LUT with a monocular depth estimator known to fail on such scenes, and compare against the paper's own depth-free YCbCr baseline (21.73 dB PSNR from its ablation). If the depth-conditioned variant fails to beat that baseline on the turbid subset, the claim that depth conditioning is the mechanism behind the improvement would be falsified.","tokens_in":15317,"feed_emoji":"🌊","tokens_out":11038,"duration_ms":97232,"temperature":0.7,"pith_summary":"DY-LUT sets out to show that the two physical facts governing underwater image degradation—propagation distance and wavelength-dependent attenuation—can be encoded directly into the architecture of an enhancement network rather than approximated by a large generic model. The paper's proposal is a compact 3.56M-parameter network whose core is a bank of learnable 4D lookup tables: the tables are queried by the luminance channel, a depth map, and two jointly learned degradation indices, then a tiny refinement module smooths the result. The authors report that this design reaches 23.245 dB PSNR on the UIEB-90 paired benchmark, wins or ties the top results on the zero-shot LSUI benchmark, and does so at 10.79 ms per image—9–304× faster than representative heavy baselines, with about 7 ms per frame at 4K using adaptive resolution. The practical stake is that real-time underwater enhancement becomes feasible on drones, ROVs, and edge hardware, and that the restored images actually help downstream detection and feature matching. What makes the argument physically grounded is the Beer–Lambert-style attenuation model: depth is not a side input but the coordinate that tells the lookup how much attenuation to reverse.","feed_headline":"Restore underwater scenes 9-304x faster via depth-aware lookup tables","feed_subtitle":"A 3.56M-parameter lookup network gets 23.2 dB PSNR at 10.8 ms and 141 fps at 4K.","key_machinery":"The engine of the method is a bank of K=3 learnable 4D lookup tables, each of size 25×25×25×25 with a three-channel YCbCr residual in the final dimension. For every pixel, quadrilinear interpolation with separable weights collects 16 neighboring table entries; the query coordinates are luminance Y, normalized depth D, and two learned degradation indices CDI1 and CDI2, predicted by a dual-branch encoder from the concatenation of YCbCr, depth, and luminance gradient. A weight branch produces image-level fusion coefficients α_k that mix the three tables, and a lightweight local-refinement network accounts for spatial consistency. Keeping the lookup 4D—rather than adding depth as a fifth axis—li","core_discovery":"On the paper's own terms, the central discovery is that depth-conditioned lookup in YCbCr space can absorb most of the work of a heavyweight restoration network. Three 25^4 lookup tables, predicted by no more than 3.56M parameters, store YCbCr residuals; each pixel queries them with coordinates (Y, D, CDI1, CDI2), where Y is luminance, D is normalized depth, and the two CDI channels are latent indices produced by a dual-branch encoder from the image and depth. Image-level fusion weights mix the three tables, and a small local-refinement module corrects edge inconsistencies. The ablation evidence shows each ingredient matters: adding depth raises PSNR by roughly 1.5–1.7 dB in both RGB and YCb","pith_inferences":["A natural extension the paper leaves implicit is confidence-aware depth conditioning: instead of treating the depth channel as equally reliable everywhere, a future variant could weight the lookup by an uncertainty estimate from the depth estimator, which would directly address the turbid-region failure mode named in the paper's limitation statement.","The non-identifiability of the two latent indices suggests they may be learning a compressed encoding of physical degradation parameters such as backscatter and attenuation coefficient; probing them on synthetic underwater scenes with known water types could reveal whether they align with those quantities.","If the depth-conditioned lookup transfers to other wavelength-dependent scattering problems—haze, murky water, atmospheric turbidity—the same 4D table structure could become a general real-time restoration primitive, using whatever depth or range sensor the platform already carries.","The modular depth interface means the same enhanced network can consume stereo disparity, lidar range, or a monocular estimate without redesign, which could simplify deployment across heterogeneous underwater vehicles."],"forward_implications":["Real-time underwater enhancement becomes feasible on constrained platforms: the same 3.56M-parameter network exceeds 141 FPS at 4K with adaptive inference and stays above 1 FPS even on a CPU-only configuration, while retaining competitive full-reference quality.","Depth does not have to come from an expensive source: the paper shows retraining for different monocular depth estimators recovers to at least 23.235 dB PSNR on UIEB-90, and switching the depth source without retraining keeps the drop within 0.763 dB.","The representation rather than the depth cue alone carries the gain: YCbCr lookup outperforms RGB lookup with and without depth, and the two learned latent indices add ~1.5 dB over fixed chrominance coordinates, suggesting the lookup structure itself is part of the explanation.","Efficiency does not appear to trade away structure needed by downstream tasks: with all else fixed, DY-LUT raises detector mAP@0.5 by 1.42 points over the strongest common baseline and increases the number of verified feature-matching correspondences by 126.","Adaptive downsampling offers a tunable speed–quality lever: at 4K it cuts latency from 145 ms to ~7 ms at the cost of 0.336 UIQM and 1.180 UISM, giving deployers a per-platform choice between maximum enhancement score and maximum throughput."],"fun_headline_variants":["Depth-aware LUTs restore underwater scenes 304x faster","Underwater 4K enhancement in 7ms via depth-aware LUTs","YCbCr beats RGB for depth-conditioned underwater enhancement","3.56M-param LUTs make underwater restoration real-time","Depth-guided lookup tables slash underwater enhancement cost"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing assumption is that the externally supplied depth map is a reliable proxy for true scene distance; in textureless or highly turbid regions the paper's own limitation statement concedes enhancement quality depends on that reliability, and if the depth is wrong the depth-conditioned lookup loses the spatial adaptivity that drives most of the quality gain.","fun_headline_variants_meta":{"raw":{"variants":["Depth-aware LUTs restore underwater scenes 304x faster","Underwater 4K enhancement in 7ms via depth-aware LUTs","YCbCr beats RGB for depth-conditioned underwater enhancement","3.56M-param LUTs make underwater restoration real-time","Depth-guided lookup tables slash underwater enhancement cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000746,"raw_usage":{"total_tokens":3186,"prompt_tokens":793,"completion_tokens":2393,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":2306}},"tokens_in":537,"tokens_out":2393,"duration_ms":17605,"temperature":1.0,"reasoning_tokens":2306,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T04:58:09.852337+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: take a benchmark of turbid or textureless underwater images, run DY-LUT with a monocular depth estimator known to fail on such scenes, and compare against the paper's own depth-free YCbCr baseline (21.73 dB PSNR from its ablation). If the depth-conditioned variant fails to beat that baseline on the turbid subset, the claim that depth conditioning is the mechanism behind the improvement would be falsified.","supporting_citations":[],"review_version":1}