{"id":"1cdec069-9053-4baa-ac92-4409c6e18bc3","arxiv_id":"2411.15740","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"LTCF-Net combines dual LAB/YUV color-space processing with transformer attention and a Fourier brightness module, reporting top PSNR/SSIM on five benchmarks with only 0.155M parameters.","lead":"This paper introduces LTCF-Net, a neural network that brightens low-light photos by processing brightness and color separately in two color spaces (LAB and YUV) and adjusting brightness in the frequency domain. The authors report higher PSNR/SSIM than many published methods while using only 0.155M parameters, but they do not release code, weights, or a detailed enough recipe to reproduce the results independently.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim is not yet supported because Table 1's comparability is unverified: the paper's SID baseline row (16.97 dB) is far below the standard SID benchmark and the RAW-to-RGB conversion is unspecified, so reported margins may be protocol artifacts.","rationale":"The reader's weakest_assumption is the correct primary concern. Table 1 is the sole quantitative support for the SOTA claim, and its validity depends on all methods being evaluated under the same conversion, splits, and metric code. The paper does not state whether baseline numbers were recomputed or copied, and no code or weights are released. The SID baseline anomaly (16.97 dB vs ~28-29 dB in the original SID paper) is a concrete signal that the evaluation protocol is not the standard SID benchmark; if the authors used a different RAW-to-RGB pipeline, all SID/SMID comparisons in Table 1 are not comparable to published results. I also note internal inconsistencies in Eq. 5 and Eq. 7, but they are secondary: even if corrected, the protocol comparability issue remains. The concrete test proposed would settle the concern by re-running the key baselines under the authors' exact pipeline; if the margins reproduce, the SOTA claim is supported, and if not, it fails. Since this matches the reader's conditional verdict, no change is needed.","tokens_in":13107,"tokens_out":8911,"duration_ms":79429,"concrete_test":"Obtain the authors' RAW-to-RGB conversion script for SID/SMID and the trained LTCF-Net weights. Re-run LTCF-Net and at least the three strongest baselines in Table 1 (Retinexformer, HVI-CIDNet, LYT-Net) with that exact conversion, the stated train/test splits, and a single PSNR/SSIM implementation. As a minimal sanity check, reproduce the SID baseline row: evaluating the official SID model under the authors' protocol should yield around 28-29 dB, otherwise the authors must show that 16.97 dB corresponds to a deliberately different protocol, in which case every baseline must be recomputed under that protocol. If LTCF-Net's margin over re-run baselines falls below the reported 0.2-0.4 dB on SID/SMID, the state-of-the-art claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that LTCF-Net outperforms SOTA on six benchmarks rests entirely on Table 1, which assumes every method was evaluated under an identical protocol: same RAW-to-RGB conversion for SID/SMID, same splits, same metric implementation. The paper does not provide code, weights, or the conversion pipeline; it only says the conversion uses 'the same in-camera signal processing as used in SID.' That is insufficient to establish comparability. Evidence that the protocol is non-standard appears inside Table 1 itself: the SID method scores 16.97 dB PSNR on the SID dataset, whereas the original SID paper reports PSNR around 28-29 dB on its Sony benchmark. A ~12 dB gap indicates that the LTCF-Net table is not measuring SID under the benchmark conditions in which published baseline numbers were produced. If baseline figures were copied from original papers or obtained with different crops, preprocessing, or test splits, the reported margins of 0.2-0.4 dB could vanish or reverse. The lack of released artifacts makes this untestable as submitted. Secondary internal inconsistencies, e.g. Eq. 5 using bounding-box coordinates {x,y,w,h} for a pixel-wise image loss and Eq. 7 giving a dimensionally incorrect LPSNR, reinforce that the text alone cannot be used to reproduce the claimed training setup.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces LTCF-Net, a low-light image enhancement network built from dual LAB/YUV color-space branches, a multi-head self-attention module, a channel denoising U-shaped block, a multi-stage squeeze-and-excitation fusion block, and a Fourier brightness processing block. The authors report state-of-the-art PSNR/SSIM on five public benchmarks (LOL-v2-real, SID, SMID, SDSD-indoor, SDSD-outdoor) with only 0.155M parameters and 10.37 GFLOPs, along with a user study and an object detection evaluation on FiveK. The central claim is that this lightweight architecture outperforms existing methods while delivering more natural color and balanced brightness.","tokens_in":13383,"tokens_out":5567,"duration_ms":42579,"significance":"If the reported results are reproducible, the paper makes a useful contribution by demonstrating that a compact design combining color-space separation, attention, and frequency-domain processing can achieve competitive or leading performance on multiple low-light benchmarks. The ablation study (Table 3) isolates the contribution of each component, which is valuable. However, the significance hinges entirely on the comparability of the baseline numbers in Table 1 and on the availability of code and weights for independent verification. As submitted, the headline SOTA claim is not verifiable because the evaluation protocol appears to differ from that used to generate published baseline results, and no code or detailed preprocessing pipeline is provided.","major_comments":[{"comment":"The comparability of the baseline results is not established. The SID row for the SID method reports a PSNR of 16.97 dB on the SID dataset, which is roughly 12 dB lower than the value reported in the original SID paper (around 28-29 dB) for its Sony benchmark. This large gap strongly suggests that the evaluation protocol used here (RAW-to-RGB conversion, image cropping, or test split) differs from the protocol under which the published baseline numbers were produced. The manuscript only states that RAW images are converted to RGB 'by applying the same in-camera signal processing as used in SID' and provides no code, no detailed pipeline, and no evidence that all baselines were recomputed under identical conditions. Since the claimed margins over the second-best methods are only 0.2-0.4 dB on most datasets, the central SOTA claim rests entirely on this unverified comparability. To support the claim, the authors must either release code/weights and a unified evaluation script that recomputes all baselines, or provide explicit, verifiable evidence that the figures in Table 1 were obtained with the same protocol for every method.","section":"Section 3.1, Table 1"},{"comment":"The smooth L1 loss is defined as a sum over i in {x, y, w, h}, which are bounding-box coordinates, not image pixels. This directly contradicts the accompanying text, which states that the loss is applied between the predicted image y_pred and the ground-truth image y_true. The definition is internally inconsistent and cannot be used to reproduce the training loss. Please correct the index set or clarify the intended pixel-wise formulation.","section":"Section 2.6, Eq. (5)"},{"comment":"The LPSNR formula is dimensionally incorrect and uses a wrong constant. For 8-bit images, PSNR equals 20*log10(255) - 10*log10(MSE), approximately 48.13 dB - 10*log10(MSE). The expression 40.0 - sum_i 20*log10(1/sqrt(MSE)) evaluates to 40.0 + 10*log10(MSE) per term, which is inverted and offset incorrectly. Moreover, the summation index and the meaning of n are undefined. Please provide a correct formula and state clearly whether LPSNR is used as a loss or as a metric.","section":"Section 2.6, Eq. (7)"},{"comment":"The user study scores and object detection AP results are reported without any measure of variance (error bars, confidence intervals, or statistical tests). Given that the claimed advantage in the user study is on the order of 0.15-0.5 points on a 0-5 scale, these results would be far more convincing with per-participant or per-image statistics and a significance test to rule out chance differences.","section":"Section 3.2, Table 2"}],"minor_comments":[{"comment":"Typo: 'We eveluate' should be 'We evaluate'.","section":"Section 3.1"},{"comment":"Typo: 'conduced' should be 'conducted'.","section":"Section 3.2"},{"comment":"References [12] and [13] are the same work (Dosovitskiy et al.), and references [42] and [43] are also the same work (Wei et al.). Please merge them or disambiguate to avoid confusion.","section":"References"},{"comment":"The caption says 'Our LTCF performs better' but the model is named LTCF-Net; please use the full name consistently.","section":"Figure 6 caption"},{"comment":"The claim that the Fourier module 'reduces noise' is supported only by a PSNR/SSIM improvement of about 1 dB and a visual example. Please consider adding a noise-specific evaluation (e.g., on synthetic noisy low-light images or with a noise estimation metric) to substantiate the qualitative claim.","section":"Section 3.4"},{"comment":"The color coding (red/yellow/blue) for best, second-best, and third-best results is not explained in the caption, and it is unclear whether LTCF-Net* is a baseline or an ablation. Please clarify the annotation.","section":"Table 1"},{"comment":"The text says 'our model registers a slightly lower PSNR on the LOL-v1 dataset,' but according to Table 1, LTCF-Net (27.07 dB) is lower than both HVI-CIDNet (27.71 dB) and LYT-Net (27.23 dB), making it third, not just slightly lower than the best. Please rephrase to accurately describe the comparison.","section":"Section 3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a preprint that currently lacks the code, weights, and preprocessing details needed to verify the central SOTA claim. The SID baseline discrepancy in Table 1 is a red flag that the evaluation protocol may be non-standard, and the editors may wish to require a reproducibility check or code release as a condition of acceptance. The loss-function errors in Eqs. (5) and (7) further underscore that the manuscript has not been carefully checked for internal consistency."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the quick read. The architecture is a reasonable engineering mashup: dual LAB/YUV branches, self-attention, a U-net denoiser, SE fusion, and a Fourier brightness block. That specific combination is new, and the model is genuinely light at 0.155M parameters. The ablation on SID shows each piece contributes, and the user study plus downstream detection test are more than most papers bother with. Credit where due: the authors are trying to evaluate on multiple fronts.\n\nThe problem is the evidence. The central SOTA claim rests on Table 1, and I can't tell if the comparison is fair. The SID baseline row reports 16.97 dB on the SID dataset; the original SID paper reports around 28-29 dB on its Sony benchmark. That 12 dB gap means either the baselines were run under a very different protocol or the numbers were copied from somewhere they don't belong. The paper says RAW images were converted to RGB using 'the same in-camera signal processing as used in SID' but gives no pipeline details. Without code, weights, or a precise description of the conversion and splits, the 0.2-0.4 dB margins over other methods are unverifiable. They could vanish or reverse under a common protocol.\n\nThere are also outright errors in the text. Eq. 5 defines smooth L1 over {x,y,w,h} — that's bounding-box regression, not a pixel loss. Eq. 7 gives a dimensional mess for the PSNR loss. These suggest the equations don't match the training code, so even the loss function is unclear. The abstract claims SOTA while Table 1 shows a loss on LOL-v1; they acknowledge this in Section 3.2, but the abstract should be consistent.\n\nWho gets value? Someone working on lightweight LLIE might find the dual color-space idea worth borrowing. But I wouldn't cite the numbers, and I wouldn't treat the SOTA claim as established. The paper needs a major revision: correct the equations, release code and weights, and run a controlled comparison where all baselines use the same conversion and splits.\n\nMy recommendation: send to a serious referee, but with the expectation of heavy revision. The architecture is plausible and the lightweight angle is worth checking; the evidence just isn't there yet.","headline":"Plausible lightweight architecture, but the SOTA claim rests on an unverifiable comparison table and some wrong equations; needs heavy revision before the numbers can be trusted.","tokens_in":13954,"tokens_out":4653,"would_cite":false,"duration_ms":41621,"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":"LTCF-Net, a dual-channel LAB/YUV transformer network with a Fourier brightness module, claims state-of-the-art results on five low-light restoration benchmarks while remaining lightweight.","keywords":["low-light image enhancement","dual color space","LAB color space","YUV color space","transformer self-attention","Fourier domain processing","lightweight network","object detection"],"falsifier":"Re-run the Table 1 experiments from released code with identical splits and the paper's RAW conversion for SID/SMID; if LTCF-Net's PSNR on any dataset drops below the leading baseline (e.g., Retinexformer or DiffLL) by more than 0.4 dB, the central superiority claim fails.","tokens_in":12860,"feed_emoji":"🌙","tokens_out":4175,"duration_ms":33592,"temperature":0.7,"pith_summary":"This paper introduces LTCF-Net, a low-light image restoration network that processes illumination and color in separate LAB and YUV channels, applies transformer self-attention to luminance, denoises chromatic components with a U-shaped block, fuses streams with squeeze-and-excitation, and adjusts brightness in the Fourier domain. The aim is to outperform current state-of-the-art methods on standard benchmarks while staying lightweight (0.155M parameters, 10.37 GFLOPs). Across LOL-v2-real, SID, SMID, SDSD-indoor, and SDSD-outdoor, the paper reports the best PSNR/SSIM among compared methods, and it shows that its enhancement improves YOLOv4-tiny object detection mean AP on FiveK to 37.7. If the comparisons hold under identical evaluation protocols, LTCF-Net offers an efficient alternative to Retinex-based and diffusion-based approaches for low-light vision.","feed_headline":"Two color spaces, one tiny net: best low-light enhancement","feed_subtitle":"With 0.155M parameters, LTCF-Net tops five low-light benchmarks and boosts night object detection.","key_machinery":"The architecture is a dual-branch network. One branch converts RGB to LAB and the other to YUV; each branch processes the luminance-like channel (L or Y) with a multi-head self-attention block plus a Fourier Brightness Processing block, while the chroma channels pass through a four-scale U-shaped Channel Denoising block. A Multi-stage Squeeze-and-Excitation Fusion block then merges luminance and denoised chroma features, and a combined pixel-plus-perceptual loss trains the whole pipeline. The Fourier block splits the frequency-domain signal into real and imaginary parts, convolves them, and re-expands them—this is claimed to remove background noise and spread brightness evenly.","core_discovery":"The paper's central claim is that separating image processing into two color spaces—LAB and YUV—lets a single-stage end-to-end network decouple illumination from chrominance more effectively than Retinex-based decomposition, and that a Fourier-domain brightness module cleans residual noise while equalizing light distribution. On its evaluation, LTCF-Net achieves 29.76 dB PSNR / 0.92 SSIM on LOL-v2-real, 26.28 / 0.70 on SID, 29.63 / 0.86 on SMID, 29.98 / 0.89 on SDSD-indoor, and 30.14 / 0.87 on SDSD-outdoor, the highest among the methods listed in Table 1, while using two orders of magnitude fewer parameters than many competitors. The paper also reports that when enhanced images are fed to YOLOv4-tiny on FiveK, its mean AP of 37.7 exceeds the other enhancement baselines.","pith_inferences":["A natural next test is a fair cross-dataset generalization check: train on SID and test on LOL-v2-real, since the paper's dual color-space separation should generalize better than Retinex-based unrolling if its decoupling claim is right.","The Fourier brightness module's noise-removal claim could be tested directly by adding synthetic Gaussian or Poisson noise to dark inputs and measuring PSNR before/after the FBP block; the paper only shows qualitative noise reduction.","The 0.155M parameter count is unusually low for a transformer-hybrid; if reproduced, it suggests the attention and Fourier modules are doing the heavy lifting rather than the convolutional denoiser.","Because the object detection test uses YOLOv4-tiny, the method's value as an image preprocessing step could differ with newer detectors; repeating the FiveK experiment with a modern detector would clarify whether the AP gain transfers."],"forward_implications":["If the reported numbers reproduce, LTCF-Net becomes a practical lightweight module for real-time low-light enhancement on mobile or embedded vision systems.","The dual color-space design suggests that processing chrominance and luminance in different color encodings is a viable alternative to Retinex-based decomposition for end-to-end training.","Improved object detection AP on FiveK indicates that low-light enhancement can serve as a preprocessing step for night-time perception stacks.","Removing the Fourier block (LTCF-Net*) gives a smaller model with only slight loss of PSNR on some datasets, offering a speed-quality trade-off.","The use of LAB and YUV together raises the question of whether other color spaces or learnable transforms would yield further gains."],"supporting_citations":[{"why":"Supplies the SID dataset and the RAW-to-RGB in-camera signal processing protocol used to generate training and test pairs.","marker":"[8]"},{"why":"Supplies the SMID benchmark dataset used for training and testing the method's generalization to smartphone-captured RAW images.","marker":"[7]"},{"why":"Supplies the SDSD static video dataset (indoor and outdoor) on which the paper reports its best PSNR/SSIM results.","marker":"[39]"},{"why":"Supplies the LOL-v1 dataset and the RetinexNet baseline architecture that the paper compares against.","marker":"[42]"},{"why":"Supplies the LOL-v2-real dataset used for the main low-light enhancement comparison.","marker":"[49]"},{"why":"Supplies the FiveK dataset and its expert-adjusted reference pairs used for the object detection downstream evaluation.","marker":"[5]"},{"why":"Provides the Retinexformer baseline, the strongest one-stage Retinex-based transformer, which LTCF-Net is designed to beat.","marker":"[6]"},{"why":"Provides the lightweight YUV transformer baseline LYT-Net, directly motivating the dual-color-space architecture.","marker":"[4]"}],"fun_headline_variants":["Tiny net, two color spaces, Fourier light: SOTA low-light","Low-light? Split LAB and YUV, add Fourier, win benchmarks","0.155M params: dual-color Fourier Transformer lifts low-light","Two spaces, one Fourier, zero Retinex: tiny SOTA","Lightweight LTCF-Net: LAB+YUV+Transformer+FFT beats SOTA"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"All state-of-the-art comparisons in Table 1 assume every baseline was evaluated under the same train/test splits, the same RAW-to-RGB conversion for SID and SMID, and the same metric code as LTCF-Net; the paper gives the conversion principle but not its exact pipeline or code, so protocol mismatches could change which method wins.","fun_headline_variants_meta":{"raw":{"variants":["Tiny net, two color spaces, Fourier light: SOTA low-light","Low-light? Split LAB and YUV, add Fourier, win benchmarks","0.155M params: dual-color Fourier Transformer lifts low-light","Two spaces, one Fourier, zero Retinex: tiny SOTA","Lightweight LTCF-Net: LAB+YUV+Transformer+FFT beats SOTA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000387,"raw_usage":{"total_tokens":2023,"prompt_tokens":907,"completion_tokens":1116,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":1029}},"tokens_in":523,"tokens_out":1116,"duration_ms":9701,"temperature":1.0,"reasoning_tokens":1029,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:57:15.962431+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the Table 1 experiments from released code with identical splits and the paper's RAW conversion for SID/SMID; if LTCF-Net's PSNR on any dataset drops below the leading baseline (e.g., Retinexformer or DiffLL) by more than 0.4 dB, the central superiority claim fails.","supporting_citations":[{"cited_title":"Seeing motion in the dark","cited_arxiv_id":null,"evidence_quote":"Supplies the SID dataset and the RAW-to-RGB in-camera signal processing protocol used to generate training and test pairs."},{"cited_title":"Learning to see in the dark","cited_arxiv_id":null,"evidence_quote":"Supplies the SMID benchmark dataset used for training and testing the method's generalization to smartphone-captured RAW images."},{"cited_title":"Seeing dynamic scene in the dark: A high-quality video dataset with mechatronic alignment","cited_arxiv_id":null,"evidence_quote":"Supplies the SDSD static video dataset (indoor and outdoor) on which the paper reports its best PSNR/SSIM results."},{"cited_title":"Sparse gradient regularized deep retinex network for robust low-light image enhancement","cited_arxiv_id":null,"evidence_quote":"Supplies the LOL-v2-real dataset used for the main low-light enhancement comparison."},{"cited_title":"Learning photographic global tonal adjustment with a database of input/output image pairs","cited_arxiv_id":null,"evidence_quote":"Supplies the FiveK dataset and its expert-adjusted reference pairs used for the object detection downstream evaluation."},{"cited_title":"Retinexformer: One-stage retinex- based transformer for low-light image enhancement","cited_arxiv_id":null,"evidence_quote":"Provides the Retinexformer baseline, the strongest one-stage Retinex-based transformer, which LTCF-Net is designed to beat."}],"review_version":1}