{"id":"e04e9bd1-e8bd-4d61-a1a8-fdf48b29f40d","arxiv_id":"2501.16211","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"UDBE trains a conditional diffusion model on synthetic brightness-shifted underwater images, then evaluates against three prior methods on UIEB, SUIM, and RUIE using PSNR, SSIM, UIQM, and UISM.","lead":"This paper presents UDBE, a diffusion model that aims to brighten underwater images by training on synthetic pairs made with random brightness changes instead of real paired photos. A smart generalist might read it because clearer underwater images would help marine robotics, inspection, and ocean science, but the current evidence is mostly based on self-created references.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table I's RUIE UIQM/UISM values contradict the claimed superiority, and the full-reference metrics only measure inversion of the synthetic brightness shift; the central claim is unsupported as printed.","rationale":"The reader's weakest assumption is the realism of the synthetic brightness shift, which is a serious external-validity problem. The most decisive issue, however, is internal: even if the synthetic benchmark were accepted as meaningful, Table I's RUIE UIQM and UISM values contradict the conclusion that UDBE outperforms the comparison methods and is robust on all four metrics. The paper's own evidence is therefore inconsistent, so the 'outperforms' claim is not merely unproven; it is contradicted by the reported numbers. If those numbers are typos, they must be corrected and the evaluation rerun; if they are not, the claim is demonstrably false on RUIE for two of the four metrics. The synthetic-shift issue compounds the problem: because the training inputs and full-reference targets are both generated from the same raw images by adding or subtracting uniform brightness values, the PSNR/SSIM columns may reflect inversion of a synthetic transformation rather than real underwater brightness enhancement. I agree with the reader's rejection verdict, but my primary load-bearing concern differs from the reader's singled-out weakest assumption; hence 'partial.' The verdict should remain as the reader gave it: REJECT, with no change needed.","tokens_in":10469,"tokens_out":6242,"duration_ms":59524,"concrete_test":"Recompute Table I's RUIE row using the released code on the same RUIE test split, independently computing UIQM and UISM from the output images. If UDBE's UIQM and UISM remain below RUIDL's values (0.829 and 4.674), Section IV-E's 'our approach outperforms the comparison methods' is false for those metrics; if the values change, the table needs correction and the evaluation should be rerun before any superiority claim is made.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that UDBE outperforms existing methods on UIEB, SUIM, and RUIE—is not supported by the paper's own quantitative evidence. In Table I, on RUIE, UDBE's UIQM is 0.4531 versus 0.829 for RUIDL and 0.624 for UDNet, and its UISM is 3.902 versus 4.674 for RUIDL and 4.005 for UDNet. Yet Section IV-E states that 'our approach outperforms the comparison methods,' and the abstract promises robustness on 'UIQM and UISM.' This is an internal contradiction: either the table or the conclusion is wrong, so no superiority claim can be drawn from the results as printed. The remaining full-reference evidence is also not probative for real underwater darkness: PSNR and SSIM are computed using high-brightness references generated by adding [50,100] to the same raw images and low-brightness inputs generated by subtracting [50,100] (Sections III-B-1 and IV-E). A model can score well by learning to invert that uniform additive shift, which does not demonstrate enhancement of wavelength-dependent attenuation, scattering, or artificial-lighting artifacts. Thus the central claim rests on an internally contradicted and externally non-representative evaluation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UDBE, an unsupervised diffusion-based method for brightness enhancement of underwater images. The method constructs synthetic high- and low-brightness versions of raw underwater images by adding or subtracting random brightness values, conditions a DDPM on color maps and SNR maps, and compares against RUIDL, UDNet, and UESAM on UIEB, SUIM, and RUIE using PSNR, SSIM, UIQM, and UISM. The authors claim that UDBE outperforms the comparison methods and is robust across the three datasets.","tokens_in":10752,"tokens_out":4767,"duration_ms":44981,"significance":"If the central claim were supported, the paper would offer a practically useful unsupervised diffusion-based alternative for underwater brightness enhancement, and the released code would aid reproducibility. The idea of conditioning a diffusion model on color and SNR maps for brightness control is appealing and the qualitative figures show some plausible improvements. However, the current evaluation does not establish the claimed superiority: the full-reference metrics are computed against synthetic references generated by the authors' own brightness operation, and the non-reference metrics on RUIE contradict the stated conclusion. The significance of the contribution therefore cannot be assessed from the evidence as printed.","major_comments":[{"comment":"The quantitative evaluation is self-referential. Both the training data and the test references are generated from the same raw images by adding or subtracting uniform random brightness values in [50,100] and [-50,-100], respectively. Section IV-E states that the generated high-brightness images are used as references for PSNR and SSIM and the low-brightness images as inputs. A model trained to invert this exact synthetic brightness shift can therefore score high on PSNR and SSIM without performing the physical operations that characterize real underwater darkness, such as wavelength-dependent attenuation, scattering, and artificial-lighting artifacts. Since UIEB provides real reference images, the authors had the option to evaluate on those; the choice of synthetic references makes the reported full-reference numbers non-representative of real enhancement.","section":"Section III-B-1 and Section IV-E"},{"comment":"The text asserts that 'our approach outperforms the comparison methods' and the abstract claims robustness on UIQM and UISM, but Table I shows UDBE's RUIE UIQM (0.4531) is below RUIDL (0.829) and UDNet (0.624), and its UISM (3.902) is below RUIDL (4.674) and UDNet (4.005). This is an internal contradiction: either the table or the conclusion is wrong, so the superiority claim cannot be accepted as printed.","section":"Table I, RUIE rows, and Section IV-E"},{"comment":"There is a mismatch between the stated reference data and the loss functions. The text says that high-brightness images H are used as reference data during the unsupervised learning process, but all loss equations compare the enhanced output B to the raw image I (Eqs. 11-15). If H is the intended target, the losses are incorrectly specified; if I is the target, the method is not optimizing toward the stated brightness-enhancement goal. This affects what the model actually learns. In addition, the method is described as unsupervised, but Section III-B/III-C constructs explicit paired synthetic data (L, H) from I and uses H as reference; the terminology should be justified or revised.","section":"Section III-C and Eqs. (11)-(15)"}],"minor_comments":[{"comment":"The low-brightness range is stated as [-50, -100]; since these values are said to be subtracted from the raw images, the sign convention makes the operation add brightness. The intended range is likely [50, 100].","section":"Section III-B-1"},{"comment":"The abstract promises an 'impressive accuracy rate' and 'robustness' on UIQM and UISM, while the conclusion says 'reasonable performance'; these characterizations should be aligned with the quantitative results.","section":"Abstract and Section V"},{"comment":"The acronym 'CLE' in 'CLE Diffusion' is not defined, and the relationship to reference [24] should be stated explicitly.","section":"Section III-C"},{"comment":"The dataset descriptions omit that SUIM is a semantic segmentation benchmark and that RUIE contains multiple evaluation subsets; the exact splits and subsets used should be reported.","section":"Section IV-A"},{"comment":"The text discusses 'PNSR' instead of 'PSNR', and the qualitative analysis would be more convincing if accompanied by quantitative brightness or illumination statistics in addition to subjective visual inspection.","section":"Section IV-E and Figure 2"},{"comment":"The angular color loss notation is unclear; the angle operation between the three-dimensional RGB vectors should be defined explicitly.","section":"Eq. (13)"}],"recommendation":"reject","confidential_remarks":"The core problem is not presentation but the experimental protocol: the central claim of superiority is contradicted by Table I and is based on a self-referential evaluation. A future revision with evaluation on real paired references, corrected loss specification, and a moderated claim could be reconsidered, but the paper as printed cannot support the abstract's claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is not ready as a research contribution, but it is not worthless. What is actually new is a specific combination: CLE Diffusion adapted to underwater brightness enhancement with color maps and SNR maps as conditioning inputs. That is an application-level novelty, not a new mechanism, and the paper honestly cites the diffusion-based underwater methods that come before it (WF-Diff, UIEDP, UW-DDPM, Tang et al.), so the 'first' claim is narrow. The writing is clear, the method description is reproducible, and the authors provide code. Credit where due: the design is coherent and the pre-processing steps are explained well enough to reimplement.\n\nThe soft spot is load-bearing. The training and evaluation are self-referential. High-brightness references H are made by adding random values in [50,100] to raw images, and low-brightness inputs L by subtracting [50,100] (Section III-B-1). PSNR and SSIM are then computed with H as the reference and L as the input (Section IV-E). That measures how well the model inverts a uniform additive brightness operation, not how well it handles real underwater darkness, which involves wavelength-dependent attenuation, scattering, and uneven artificial light. So the full-reference numbers are not probative.\n\nThe no-reference metrics make the problem worse. On RUIE, Table I shows UDBE's UIQM at 0.4531 versus RUIDL's 0.829, and UISM at 3.902 versus RUIDL's 4.674. The text says 'our approach outperforms the comparison methods' and the abstract promises robustness on UIQM and UISM. That is an internal contradiction you can see at a glance. Either the table or the conclusion is wrong; the superiority claim cannot stand as printed.\n\nThere are also smaller issues: the cited diffusion baselines are not included in the comparison, and the five loss weights are hand-tuned with no sensitivity analysis. None of that would be fatal on its own, but combined with the evaluation problem it sinks the current version.\n\nWho gets value from this paper? Someone working on underwater image enhancement might read it as a cautionary example of evaluation design, and the conditioning idea could be worth borrowing. But the paper does not demonstrate that UDBE beats existing methods in any realistic sense.\n\nRecommendation: send it to peer review rather than desk-rejecting, because the method is concrete, the code is promised, and the flaws are fixable with a real evaluation protocol. But as a referee I would ask for major revision and would not accept it in its current form.","headline":"The method is a reasonable adaptation of CLE Diffusion to underwater brightness, but the evaluation only measures inversion of the authors' own synthetic brightness shift, and Table I contradicts the text on RUIE, so the central claim is unsupported.","tokens_in":11289,"tokens_out":1942,"would_cite":false,"duration_ms":19689,"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":"UDBE uses an unsupervised conditional diffusion process to brighten underwater images without paired references.","keywords":["underwater image enhancement","brightness enhancement","diffusion model","unsupervised learning","conditional diffusion","DDPM","image quality assessment","SNR map"],"falsifier":"Collect a set of real underwater scenes photographed in darkness and then re-photographed under controlled artificial lighting to form true bright references; run UDBE on the dark versions and compute PSNR/SSIM against those true references. If the scores are no better than a simple histogram stretch or than inverting the uniform shift, the synthetic training assumption is the limiting factor; if they are close to the paper's reported values, the assumption transfers.","tokens_in":10287,"feed_emoji":"🌊","tokens_out":14025,"duration_ms":118745,"temperature":0.7,"pith_summary":"The paper proposes UDBE, an unsupervised diffusion method for increasing brightness in underwater images without paired reference images. It creates synthetic training pairs by subtracting or adding random brightness values in the 50–100 range to raw underwater images, then trains a DDPM-based network to reverse the darkening. Color maps and signal-to-noise maps are appended to the input so the model brightens without distorting color. On the UIEB, SUIM, and RUIE benchmarks, the paper reports higher PSNR, SSIM, UIQM, and UISM scores than three prior enhancement methods, alongside visually clearer output. If correct, this would make diffusion-based brightness enhancement practical for underwater datasets that lack ground-truth bright versions.","feed_headline":"Unsupervised diffusion model brightens underwater images","feed_subtitle":"Trained on synthetic brightness shifts, it beats three prior methods on underwater benchmarks while keeping color.","key_machinery":"The central mechanism is the conditional reverse diffusion of a DDPM, steered by three conditioning signals. A FiLM layer (feature-wise linear modulation) in every U-Net layer applies an affine transformation driven by the desired brightness level; the network input also concatenates a color map, in which each RGB channel is normalized by its maximum value, and an SNR map, defined as the Gaussian-blurred image divided by the absolute difference between the image and its blur, which highlights dark and noisy regions. These signals let one network brighten an image while keeping colors stable. The training objective is a weighted sum of LPIPS, SSIM, MSE, brightness, and color losses, with the brightness and color terms added after the first twenty epochs.","core_discovery":"The central claim is that brightness degradation in underwater images can be learned and reversed in a fully unsupervised way by a conditional diffusion model. For each raw image, the method constructs a low-brightness input by subtracting a random value between 50 and 100 and a high-brightness reference by adding a random value between 50 and 100. The network, a U-Net inside a DDPM with FiLM brightness conditioning, is trained with a weighted mix of LPIPS, SSIM, MSE, grayscale brightness, and angular color losses. During inference, the raw dark image is concatenated with a normalized color map and an SNR map, and the reverse diffusion process produces the brightened result. The paper reports that UDBE surpasses RUIDL, UDNet, and UESAM in PSNR and SSIM on all three datasets, and in most UIQM and UISM comparisons, and it presents this as the first diffusion-based brightness enhancement for underwater images.","pith_inferences":["The reported full-reference scores measure how well the model inverts a uniform additive brightening of the same image, so they should not be read as direct evidence about real depth-induced attenuation; testing on true paired underwater captures is the natural next step.","The paper's own table shows that on RUIE the non-reference metrics UIQM and UISM are lower for UDBE than for RUIDL, so the 'outperforms' claim is strongest for full-reference similarity on synthetic brightness shifts and is not uniform across every metric.","A testable extension is to replace the uniform brightness augmentation with a wavelength-dependent attenuation and scattering model; the change in benchmark scores would separate what UDBE learns about brightness inversion from what it learns about real underwater physics.","The color-map/SNR-map conditioning is a general mechanism that could transfer to other unpaired image restoration problems, such as low-light photography or turbid-water video, where color and noise are the main degradation channels."],"forward_implications":["If the claim is correct, underwater enhancement no longer requires paired reference images, so any raw underwater collection can serve as training data.","The same conditioning recipe—brightness level, color map, and SNR map—could be reused for other unpaired restoration tasks where color and noise structure are known.","Diffusion-based restoration would become competitive with established CNN and self-supervised baselines on standard underwater benchmarks, at least for brightness.","Because inference uses DDIM sampling, the enhanced output is produced without the full iterative cost of a standard diffusion sampler."],"supporting_citations":[{"why":"Provides the DDPM forward and reverse diffusion process and the simplified denoising loss that UDBE is built on.","marker":"[3]"},{"why":"Provides the RUIDL enhancement method that UDBE is compared against.","marker":"[5]"},{"why":"Provides the UDNet unsupervised enhancement baseline used in the comparison.","marker":"[18]"},{"why":"Provides the UESAM self-supervised enhancement baseline used in the comparison.","marker":"[19]"},{"why":"Source of the controllable light enhancement diffusion idea and the brightness loss that UDBE adapts to the underwater domain.","marker":"[24]"},{"why":"Supplies the FiLM layer used to condition each U-Net layer on the brightness level.","marker":"[25]"},{"why":"Provides the UIEB underwater image benchmark used for training and quantitative evaluation.","marker":"[30]"},{"why":"Provides the SUIM underwater image dataset used for training and evaluation.","marker":"[31]"},{"why":"Provides the RUIE real-world underwater enhancement benchmark used for training and evaluation.","marker":"[32]"},{"why":"Defines the UIQM and UISM non-reference underwater image quality metrics used in the experiments.","marker":"[34]"}],"fun_headline_variants":["Unsupervised diffusion model brightens dark underwater scenes","UDBE: First diffusion method to brighten images without labels","Brightening deep-sea shots without any supervision","Diffusion lightens dark underwater photos, no labels needed"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that dark underwater images can be simulated by subtracting a uniform brightness value between 50 and 100 from an already captured raw image, and that adding 50 to 100 produces a correct bright reference; both training and evaluation depend on that synthetic brightness shift matching real underwater darkness.","fun_headline_variants_meta":{"raw":{"variants":["Unsupervised diffusion model brightens dark underwater scenes","UDBE: First diffusion method to brighten images without labels","Brightening deep-sea shots without any supervision","Diffusion lightens dark underwater photos, no labels needed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000709,"raw_usage":{"total_tokens":3191,"prompt_tokens":941,"completion_tokens":2250,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":2185}},"tokens_in":557,"tokens_out":2250,"duration_ms":16542,"temperature":1.0,"reasoning_tokens":2185,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T13:35:55.730832+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect a set of real underwater scenes photographed in darkness and then re-photographed under controlled artificial lighting to form true bright references; run UDBE on the dark versions and compute PSNR/SSIM against those true references. If the scores are no better than a simple histogram stretch or than inverting the uniform shift, the synthetic training assumption is the limiting factor; if they are close to the paper's reported values, the assumption transfers.","supporting_citations":[{"cited_title":"Diving into clarity: Restoring underwater images using deep learning,","cited_arxiv_id":null,"evidence_quote":"Provides the RUIDL enhancement method that UDBE is compared against."},{"cited_title":"Adaptive uncertainty distribution in deep learning for unsupervised underwater image enhancement,","cited_arxiv_id":null,"evidence_quote":"Provides the UDNet unsupervised enhancement baseline used in the comparison."},{"cited_title":"Underwater enhancement based on a self-learning strategy and attention mechanism for high-intensity regions,","cited_arxiv_id":null,"evidence_quote":"Provides the UESAM self-supervised enhancement baseline used in the comparison."},{"cited_title":"Cle diffusion: Controllable light enhancement diffusion model,","cited_arxiv_id":null,"evidence_quote":"Source of the controllable light enhancement diffusion idea and the brightness loss that UDBE adapts to the underwater domain."},{"cited_title":"An underwater image enhancement benchmark dataset and beyond,","cited_arxiv_id":null,"evidence_quote":"Provides the UIEB underwater image benchmark used for training and quantitative evaluation."},{"cited_title":"Semantic Segmentation of Underwater Imagery: Dataset and Benchmark,","cited_arxiv_id":null,"evidence_quote":"Provides the SUIM underwater image dataset used for training and evaluation."},{"cited_title":"Real-world underwater enhancement: Challenges, benchmarks, and solutions under natural light,","cited_arxiv_id":null,"evidence_quote":"Provides the RUIE real-world underwater enhancement benchmark used for training and evaluation."},{"cited_title":"Human-visual-system-inspired underwater image quality measures,","cited_arxiv_id":null,"evidence_quote":"Defines the UIQM and UISM non-reference underwater image quality metrics used in the experiments."}],"review_version":1}