{"id":"cb681fd8-b035-4140-9616-e7e6270ae2fc","arxiv_id":"2501.12235","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DLEN reports state-of-the-art PSNR/SSIM on LOL benchmarks using a dual-branch transformer with a learnable wavelet illumination module.","lead":"DLEN is a neural network that brightens low-light photos using two attention branches, one guided by illumination and one that repairs image structure, plus a learnable wavelet stage to keep edges and textures sharp. The paper reports higher PSNR and SSIM scores than several published methods on the LOL v1 and v2 benchmarks, and code is promised on GitHub.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'outperforms SOTA' claim rests on a 0.01 dB LOLv1 margin over Retinexformer with no error bars, and Table II shows an ablation that beats the full model; reproducibility is the load-bearing gap.","rationale":"The paper is incremental but architecturally plausible, and no formal proof or verified code is provided. The Reader's weakest assumption about wavelet filterbank validity is legitimate: Section III-C treats h0 and h1 as a wavelet pair in Eqs. (19) to (22), but no orthogonality or reconstruction constraint is imposed. However, that assumption affects the novelty claim; it cannot by itself falsify the reported empirical scores. The empirical SOTA claim is more directly threatened by the absence of any variability estimate and by the tiny margins. On LOLv1, 23.942 versus 23.932 dB is a 0.0004 relative difference on 15 test images, and standard benchmark variation across seeds and training runs is typically larger than this. On LOLv2-real the margin is 1.0 dB, but Table II shows the supposedly beneficial LWN hurts PSNR by 0.063 dB on that same dataset, which suggests the reported numbers are not stable enough to support the strong abstract claim. A simple multi-seed rerun would settle this. I therefore keep the Reader's CONDITIONAL verdict rather than escalating to REJECT: the architecture is coherent, and the reported numbers, if reproducible, would constitute a modest advance. The concern could be remedied by adding error bars, more current baselines, and a corrected or re-run ablation.","tokens_in":12377,"tokens_out":6208,"duration_ms":60308,"concrete_test":"Run the released or independently reimplemented DLEN and Retinexformer training pipelines on LOLv1 and LOLv2-real with at least five random seeds each, at the reported 128x128 setting, and report per-seed PSNR/SSIM, mean plus/minus standard deviation, and a paired test over the 15 and 100 test images. If the LOLv1 DLEN minus Retinexformer gap does not exceed seed-to-seed variation (for example, less than 0.05 dB), or if the LOLv2-real w/o-LWN ablation still beats the full model, the abstract SOTA claim and the LWN contribution claim are not sustained. A secondary check: print the learned h0 and h1 filters and verify the perfect-reconstruction or orthogonality condition used in Eqs. (19) to (22).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claim is that DLEN outperforms state-of-the-art methods. The only quantitative support is Table I, which reports DLEN at 23.942 dB PSNR on LOLv1 versus 23.932 dB for Retinexformer — a 0.01 dB gap on a 15-image test set. No error bars, multiple seeds, or statistical significance tests are given, so the LOLv1 margin is within ordinary training noise; the same concern applies to the 0.01 SSIM difference (0.841 vs 0.831). Table II adds internal evidence of instability: removing the Learnable Wavelet Network improves LOLv2-real PSNR from 22.230 to 22.293, directly contradicting the claim that the full setup achieves the highest PSNR and SSIM values. These two facts together make the load-bearing 'outperforms SOTA' assertion unsupported as reported. The unconstrained learned wavelet filters identified by the Reader are a real novelty concern, but they are secondary: even if h0 and h1 are arbitrary convolutions, the empirical claim could still hold; the reverse is not true.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes DLEN, a low-light image enhancement network that combines a Light Component Predictor with a learnable wavelet module and a Dual-Branch Restorer containing two attention mechanisms (MIAB and SEAB). The paper claims state-of-the-art performance on the LOLv1 and LOLv2-real benchmarks based on PSNR and SSIM comparisons in Table I, and it presents ablation studies in Table II as well as qualitative results. The central claim, stated in the abstract and introduction, is that DLEN outperforms existing methods through extensive experiments.","tokens_in":12704,"tokens_out":8568,"duration_ms":76153,"significance":"If the reported results are reproducible and the internal inconsistency in the ablation study is resolved, DLEN would be a competitive architecture for low-light image enhancement, although its measured margin over Retinexformer on LOLv1 is only 0.01 dB and it trails MIRNet in PSNR on that dataset. The combination of illumination-guided attention with a frequency-domain module is a plausible and interesting direction. The paper gives a clear architectural description and provides a code link, which are strengths; however, the evidence as presented is not sufficient to establish the claimed superiority, and one ablation row directly contradicts the main text.","major_comments":[{"comment":"The claim that DLEN 'outperforms state-of-the-art methods on standard benchmarks' is not supported by Table I as reported. On LOLv1, MIRNet achieves a higher PSNR (24.140 dB) than DLEN (23.942 dB), and the margin over Retinexformer is only 0.01 dB PSNR and 0.01 SSIM on a 15-image test set, with no error bars, repeated runs, or statistical significance tests. Section IV-C reports only average gains across the two datasets, which conceals the LOLv1 deficit relative to MIRNet. The paper should report per-dataset and per-baseline results with uncertainty quantification, and the abstract and contribution claims should be qualified accordingly.","section":"Table I and Section IV-C"},{"comment":"The ablation study contradicts the main text: removing the Learnable Wavelet Network improves LOLv2-real PSNR from 22.230 to 22.293 dB, while Section IV-D states that 'our full setup achieves the highest PSNR and SSIM values.' This internal inconsistency must be addressed, either by explaining the variation (for example, training noise) or by re-running the ablation, and it weakens the causal claim that the LWN is beneficial as claimed.","section":"Table II and Section IV-D"},{"comment":"The simplification from Eq. (3) to Eq. (4), written as I_lu = I ⊙ L~ = R + C, is not derived, and it is not evident that the left-hand side follows from Eq. (3) for any definition of L~ given in the text. The manuscript should provide the missing derivation or restate the equation with an explicit definition of C; as written, this step is a gap in the formulation of the network in Eqs. (5)-(7).","section":"Section III-A, Eq. (4)"},{"comment":"The learnable filters h0 and h1 are used to define a wavelet filter bank, but the paper imposes no orthogonality, reconstruction, or normalization constraint on them, so after training there is no guarantee that they separate low- and high-frequency components as claimed. This undermines the frequency-domain rationale for the LWN as a 'learnable wavelet transform.' The authors should either add constraints or demonstrate empirically that the learned filters behave as a filter bank, for example by visualizing their frequency responses or the resulting subbands.","section":"Section III-C, Eqs. (19)-(22)"}],"minor_comments":[{"comment":"The symbols ILB and SEB are introduced in Eq. (6), but the text later uses SEAB for the structure enhancement branch; a single consistent naming convention should be used throughout.","section":"Eq. (6) and Section III-B"},{"comment":"Several entries in the reference list, including [38] through [53], do not appear to be cited in the body of the paper, and some citations such as [7] omit co-authors; the bibliography should be reconciled with the in-text citations.","section":"References"},{"comment":"The maximum pixel value is denoted R in the PSNR formula, which collides with the reflection image R defined in Eq. (1); a different symbol such as P_max would avoid the ambiguity.","section":"Section IV-B"},{"comment":"The implementation section reports image resolution and batch sizes but not the total number of parameters, FLOPs, or training epochs; including these would allow readers to assess the computational overhead of the dual-branch architecture.","section":"Section IV-A"}],"recommendation":"major_revision","confidential_remarks":"The reference list includes a large block of entries with no in-text citations, which is a completeness issue that should be checked by the editor. The code repository link lacks a commit hash or version tag, so the reported numbers cannot be reproduced at present. The internal contradiction between Table II and Section IV-D is the most serious issue and should be resolved before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a classic incremental architecture paper for low-light enhancement, and the central 'outperforms SOTA' claim does not survive contact with the tables. On LOLv1 the gain over Retinexformer is 0.01 dB PSNR on a 15-image test set; that is training noise. The one real bright spot is LOLv2-real, where DLEN reports 22.230 vs 21.230 for Retinexformer, but there are no error bars, so we can't tell if that is meaningful.\n\nWhat is actually new is the specific combination: a learnable wavelet module lifted from deblurring [31], an illumination-guided attention block, and a structure branch, assembled on top of Retinexformer. Each piece is known, but the combination is theirs, and they describe it clearly. They also provide ablation studies for the two main components and qualitative figures. Credit where due: the related work is fair, and the overlapping citation to RetinexMamba appears only in passing and does not support the headline result.\n\nSoft spots, in order of severity. First, the ablation contradiction: Table II shows 'w/o LWN' at 22.293 PSNR on LOLv2-real, which is higher than the full model's 22.230. The text states the full setup achieves the highest values. That is an internal inconsistency and needs explanation or correction. Second, the evaluation standards: no multiple seeds, no significance tests, no details on epochs or learning rate, and the code link has no commit hash. Third, Eq. (4) is asserted rather than derived, and the perturbation term C is undefined; this is a motivation gap rather than a fatal one. Fourth, the 'learnable wavelet' places no orthogonality or reconstruction constraint on h0 and h1, so after training the filters may be arbitrary convolutions. That undercuts the frequency-domain narrative, though it would not by itself invalidate the empirical network.\n\nOverall, the architecture is plausible and the paper is honest in its coverage of the literature. The flaws are fixable, not structural. The paper would benefit from a rigorous round of review that demands better statistics and a resolution of the ablation mismatch.\n\nWho gets value: readers tracking Retinexformer variants and wavelet-based image restoration. It deserves a serious referee; I would not desk-reject it, but I would expect major revision.","headline":"A plausible incremental Retinexformer variant whose SOTA claim rests on a 0.01 dB edge and is undercut by an ablation table that beats the full model on one dataset.","tokens_in":13200,"tokens_out":4876,"would_cite":false,"duration_ms":42865,"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":"A wavelet-aided, dual-branch transformer network is claimed to set new state-of-the-art scores on low-light image enhancement benchmarks.","keywords":["low-light image enhancement","learnable wavelet transform","Retinex theory","Transformer attention","illumination estimation","dual-branch restorer","LOL dataset","PSNR and SSIM"],"falsifier":"Take the trained filters $\\vec{h}_0$ and $\\vec{h}_1$ and compute the four subband outputs on an image with smooth regions; if the high-frequency subbands are not near zero on those regions, or if the pair fails a quadrature-mirror condition such as $h_1[n] = (-1)^n h_0[N-1-n]$, then the module is not performing a wavelet separation and the frequency-domain rationale is unsupported.","tokens_in":12167,"feed_emoji":"💡","tokens_out":9459,"duration_ms":80559,"temperature":0.7,"pith_summary":"Low-light images lose brightness, contrast, and fine detail, and the paper argues that the fix should happen in both the spatial and frequency domains at once. It proposes DLEN, a network whose illumination-estimation stage contains a learnable wavelet transform that separates high- and low-frequency features, followed by a dual-branch Transformer restorer: one branch uses illumination-guided attention to brighten dark regions, the other repairs structure. On the standard LOL benchmarks, DLEN reports 23.942 dB PSNR / 0.841 SSIM on LOLv1 and 22.230 dB PSNR / 0.854 SSIM on LOLv2-real, the highest figures among the methods compared, with a 0.01 dB lead over the best prior one-stage Retinex transformer on LOLv1 and about a 1 dB lead on LOLv2-real. If the claim stands, it means combining frequency-domain separation with illumination-aware attention is a practical recipe for preserving texture while correcting exposure.","feed_headline":"Wavelet-aided transformer tops low-light image benchmarks","feed_subtitle":"DLEN reports 23.94 dB PSNR on LOLv1 and 22.23 dB on LOLv2-real, above every compared baseline.","key_machinery":"The load-bearing object is the Learnable Wavelet Module (LWN), a trainable 1D filter pair $\\vec{h}_0$ and $\\vec{h}_1$ whose outer products form the four 2D subband filters $G_{ll}$, $G_{lh}$, $G_{hl}$, and $G_{hh}$, so a feature map is split into low- and high-frequency components before restoration. Two attention blocks carry the dual-branch restorer: the Multi-Illumination Attention Block (MIAB), where the illumination feature $Y$ scales the values before softmax attention, and the Structure Enhancement Attention Block (SEAB), where pointwise and depthwise convolutions enrich queries, keys, and values and the illumination map gates the output. Together these pieces turn the Retinex decomposition and wavelet subbands into residual corrections that preserve texture and structure.","core_discovery":"On the paper's own terms, the discovery is that a low-light enhancement network can be made state-of-the-art by inserting a learnable discrete wavelet module into the illumination estimator and splitting the restorer into two Transformer-based branches. The Light Component Predictor decomposes the input under a perturbed Retinex model $I = (R + \\tilde{R}) \\odot (L + \\tilde{L})$, produces an illuminated image $I^{lu}$ and an illumination feature $F^{lu}$; the learnable wavelet then splits $F^{lu}$ into low- and high-frequency subbands before the restorer. The restorer's illumination branch modulates self-attention with the illumination feature via $(Y \\odot V)\\,\\mathrm{softmax}(K^{T}Q/\\alpha)$, while the structure branch applies depthwise-convolution-enriched attention and multiplies by the illumination map. Combined, the branches output residuals added to $I^{lu}$, and the network is trained end-to-end with mean absolute error. As reported in Table I, the result is the top PSNR/SSIM among all compared supervised methods on both LOLv1 and LOLv2-real.","pith_inferences":["Editorial inference: the learned filters are never constrained to satisfy orthogonality or perfect-reconstruction conditions, so at convergence they may act as ordinary separable convolutions; if so, the LWN's benefit would come from extra capacity rather than true frequency separation.","Editorial inference: a direct test would replace the learned filter pair with fixed Haar filters and with free $1 \\times 1$ convolution kernels while holding everything else fixed; the paper's reasoning predicts the learned pair should beat both, but this experiment is not run.","Editorial inference: the same dual-branch, prior-guided attention design could transfer to other restoration tasks with spatially varying conditions, such as underwater enhancement or shadow removal, where a prior map plays the role of the illumination feature."],"forward_implications":["If the reported scores hold, DLEN becomes the top-scoring compared method on LOLv1 and LOLv2-real, so downstream tasks such as detection and recognition on enhanced images inherit a modest quantitative improvement.","The ablation numbers attribute roughly 0.48 dB on LOLv1 to the learnable wavelet module and roughly 1.1 dB to the structure branch, implying each component carries part of the gain.","Because the learnable wavelet is a frequency-domain front end, it can be attached to other Retinex-based illumination estimators without redesigning the restorer.","Training with only mean absolute error and standard augmentation is enough to reach these scores, suggesting the architecture, not a bespoke loss, is doing the work."],"supporting_citations":[{"why":"Supplies the Retinex decomposition $I = R \\odot L$ on which the Light Component Predictor is built.","marker":"[4]"},{"why":"The one-stage Retinex-based transformer architecture that DLEN extends with the learnable wavelet and dual-branch restorer, and the strongest supervised baseline on LOLv1.","marker":"[6]"},{"why":"Provides the LOLv1 paired training and test data used for evaluation.","marker":"[10]"},{"why":"Provides the LOLv2-real and LOLv2-synthetic paired data used for evaluation.","marker":"[11]"},{"why":"Supplies the learnable discrete wavelet transform design that the LWN module adapts to low-light enhancement.","marker":"[31]"},{"why":"Retinex-based deep unfolding network used as a supervised comparison baseline in Table I.","marker":"[22]"}],"fun_headline_variants":["Dual-branch transformer with learnable wavelet tops low-light benchmarks","Wavelet-powered dual-branch transformer boosts low-light image enhancement","Light enhancement via dual-domain transformer and learnable wavelets","Transformer pair with wavelet module sets new low-light enhancement record","DLEN: wavelet-aided transformer excels in low-light image enhancement"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The frequency-domain story rests on the assumption that the two learned 1D filters still behave as a genuine low-pass and high-pass wavelet pair after training, but nothing in the training procedure forces them to.","fun_headline_variants_meta":{"raw":{"variants":["Dual-branch transformer with learnable wavelet tops low-light benchmarks","Wavelet-powered dual-branch transformer boosts low-light image enhancement","Light enhancement via dual-domain transformer and learnable wavelets","Transformer pair with wavelet module sets new low-light enhancement record","DLEN: wavelet-aided transformer excels in low-light image enhancement"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000723,"raw_usage":{"total_tokens":3273,"prompt_tokens":1006,"completion_tokens":2267,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":622,"completion_tokens_details":{"reasoning_tokens":2181}},"tokens_in":622,"tokens_out":2267,"duration_ms":16958,"temperature":1.0,"reasoning_tokens":2181,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:22:41.768475+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained filters $\\vec{h}_0$ and $\\vec{h}_1$ and compute the four subband outputs on an image with smooth regions; if the high-frequency subbands are not near zero on those regions, or if the pair fails a quadrature-mirror condition such as $h_1[n] = (-1)^n h_0[N-1-n]$, then the module is not performing a wavelet separation and the frequency-domain rationale is unsupported.","supporting_citations":[{"cited_title":"Lightness and retinex theory,","cited_arxiv_id":null,"evidence_quote":"Supplies the Retinex decomposition $I = R \\odot L$ on which the Light Component Predictor is built."},{"cited_title":"Retinex- former: One-stage retinex-based transformer for low-light image en- hancement,","cited_arxiv_id":null,"evidence_quote":"The one-stage Retinex-based transformer architecture that DLEN extends with the learnable wavelet and dual-branch restorer, and the strongest supervised baseline on LOLv1."},{"cited_title":"Sparse gradient reg- ularized deep retinex network for robust low-light image enhancement,","cited_arxiv_id":null,"evidence_quote":"Provides the LOLv2-real and LOLv2-synthetic paired data used for evaluation."},{"cited_title":"Convo- lutional and fourier neural networks for speckle denoising of wrapped phase in digital holographic interferometry,","cited_arxiv_id":null,"evidence_quote":"Supplies the learnable discrete wavelet transform design that the LWN module adapts to low-light enhancement."},{"cited_title":"Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement,","cited_arxiv_id":null,"evidence_quote":"Retinex-based deep unfolding network used as a supervised comparison baseline in Table I."}],"review_version":1}