{"id":"3e4b75a9-a635-4d5a-a8da-cfd90af6b337","arxiv_id":"2508.02168","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"A new paired dataset and Retinex-based network, RLN2, for restoring images captured under multiple colored light sources to ambient-normalized versions.","lead":"This paper introduces CL3AN, a new high-resolution dataset of photos taken under multiple colored lights, each paired with a uniformly lit reference image. It also presents RLN2, a Retinex-inspired network that aims to restore such photos to ambient lighting with fewer artifacts than existing methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"CL3AN's paired ground truth is the load-bearing assumption: if colored-light and ambient captures are not pixel-aligned with matched exposure and camera settings, every benchmark comparison loses meaning; the excerpt does not document this.","rationale":"The reader's weakest assumption is exactly the load-bearing point: the validity of the CL3AN paired-capture setup. I agree with that identification and sharpen it to include matched camera settings and processing, not only geometric alignment. The colored-light input and ambient reference must be produced under identical acquisition parameters; otherwise the network is trained to invert a mapping that is not physically well-defined, and the claimed comparisons are uninterpretable. The supplied text provides Figure 2(D) and a verbal description of the setup, but no capture, alignment, or post-processing details, and the experimental sections and tables are missing. This concern cannot be resolved from the excerpt alone, so the honest status remains unverdictable. Because my read does not move the reader's verdict, the recommendation is unchanged: the paper should be marked UNVERDICTED pending access to dataset construction details and experiments. I am not treating the absence of the experiments as evidence of failure; it is simply an evidence gap, and the repository release is a concrete way to close it.","tokens_in":8825,"tokens_out":4633,"duration_ms":59063,"concrete_test":"Inspect the released RLN2 repository (github.com/fvasluianu97/RLN2) for the CL3AN capture protocol and for EXIF or processing metadata on a sample of pairs. Compute dense optical flow between colored-light and ambient images in regions not directly lit by colored sources, and compare EXIF exposure, ISO, aperture, and white-balance settings between the two captures. If median flow exceeds one pixel, or if exposure or white-balance settings differ between members of any pair, the pixel-aligned equal-capture assumption fails and the benchmark ground truth is compromised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that RLN2 outperforms on CL3AN depends entirely on the validity of the CL3AN paired captures. In Figure 2(D), the proposed setup uses RGB direct lighting and an ambient-lit reference, but the excerpt supplies no capture protocol: no statement that the two passes share the same camera, exposure, aperture, white-balance, or tone-mapping, and no alignment procedure for the paired images. This is not merely a documentation gap. The 'ambient-normalized' reference is well-defined only if the ambient image is captured under the same static scene and the same camera settings, and if the colored-light pass does not introduce effects such as specular highlights, interreflections, or sensor saturation that the ambient pass cannot represent as a target. The text concedes that the predecessor AMBIENT6K is limited by the coupling of direct and ambient lighting parameters; CL3AN must demonstrate that dropping the color-consistency constraint does not also break exposure or geometry consistency. Since the experimental section is absent, this cannot currently be checked from the manuscript.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes CL3AN, described as the first large-scale, high-resolution dataset for Ambient Lighting Normalization (ALN) under multiple colored light sources, together with RLN2, a learning framework that performs illumination-reflectance decomposition guided by explicit chromaticity-luminance components. The paper argues that existing datasets and methods oversimplify illumination by assuming single or white-balanced light sources, and that CL3AN addresses this gap by using RGB direct lighting without color-consistency constraints. The provided text contains the abstract, introduction, the beginning of Section 3, and references, but no experimental section, dataset statistics, evaluation tables, or complete method details. The central claims are therefore not verifiable from the manuscript as presented.","tokens_in":9017,"tokens_out":3421,"duration_ms":45174,"significance":"If the claims hold, CL3AN would be a valuable new benchmark for a realistic and under-addressed problem, and RLN2 would offer a competitive solution at moderate computational cost. The paper identifies a genuine limitation of prior datasets such as AMBIENT6K and proposes a plausible direction via colored-light direct lighting. The stated intention to release code, models, and benchmark data is a positive contribution. However, the current manuscript provides no evidence for the headline claims: no dataset statistics, no capture validation, no evaluation protocol, no results, and no complete method description. The significance cannot be assessed until these are supplied.","major_comments":[{"comment":"The load-bearing assumption of CL3AN is that the colored-light input and the ambient-lit reference are pixel-aligned captures of the same scene under identical camera settings. The text states that the direct lighting setup 'is based on RGB lights, dropping the color consistency constraint,' but it does not specify the camera, exposure, aperture, white balance, tone mapping, or any alignment procedure used to obtain the paired images. Nor does it address effects such as specular highlights, interreflections, or sensor saturation that the ambient reference cannot represent. Without this capture protocol and alignment validation, the ground truth for the benchmark is not well-defined and every downstream comparison loses meaning.","section":"Section 1, Figure 2(D)"},{"comment":"The abstract claims 'Extensive evaluations on existing benchmarks and our dataset demonstrate the effectiveness of our approach' and 'highly competitive computational cost,' but the supplied text contains no experimental section, no tables, no metrics, no dataset statistics, no ablations, and no evaluation protocol. The paper cannot be assessed for soundness until these are provided. Please include dataset size and resolution, number of scenes and lighting configurations, train/test splits, evaluation metrics, comparison methods, and runtime or FLOPs measurements.","section":"Abstract and full text"},{"comment":"Section 3 begins with the sentence 'The core of our work is extending the study of Ambient Lighting Normalization to direct color lighting' and then jumps to a figure caption and an incomplete paragraph starting 'Provided statistics, such as.' The actual RLN2 architecture, the 'explicit chromaticity-luminance components guidance,' the loss functions, and the training details are absent. It is therefore impossible to evaluate the novelty of the method, to verify that the claimed decomposition is actually learned, or to reproduce the approach. Please provide a complete method section with equations, a network diagram, and training hyperparameters.","section":"Section 3"},{"comment":"The claim that CL3AN is 'the first large-scale, high-resolution dataset of its kind' is unsupported by any concrete numbers. The text does not define what 'large-scale' and 'high-resolution' mean in this context, nor does it compare the dataset size, resolution, or diversity with existing benchmarks such as ISTD/ISTD+, WSRD, AMBIENT6K, and LSMI. Additionally, since WSRD and AMBIENT6K are from the same group as the current paper, the relationship and independence of the new benchmark should be clarified when reporting comparisons on those datasets.","section":"Abstract and Figure 2"}],"minor_comments":[{"comment":"The manuscript is missing Section 2 (related work is apparently absent) and the numbering jumps from Section 1 to Section 3; this should be fixed in a complete version.","section":"Overall structure"},{"comment":"The paragraph beginning 'Provided statistics, such as' is an incomplete sentence and the statistics it refers to are not defined; please complete the thought or remove the fragment.","section":"Page 4 paragraph"},{"comment":"The affiliation contains 'W¨urzburg' with an umlaut encoding error; it should be typeset as 'Würzburg.'","section":"Author affiliation"},{"comment":"The four subfigures (A)-(D) are described in the caption, but the text does not consistently refer to all of them; please add explicit cross-references in the text.","section":"Figure 2"},{"comment":"Reference [67] is cited as 'A Vaswani' with an incomplete author list; please use the full 'Vaswani et al.' citation for 'Attention is all you need.'","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The submitted text appears to be missing its experimental section and a substantial part of Section 3; please verify that the file was not truncated in submission. In addition, the independence of the new benchmark relative to prior datasets from the same group (WSRD, AMBIENT6K) should be explicitly addressed in the final version, since the claimed comparisons on those benchmarks may not be considered fully independent."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know: this paper has a real gap to fill—a large-scale dataset of images under multiple colored lights paired with ambient-lit references—but the version we saw is missing every piece of evidence that would make the dataset trustworthy. The stress-test note is right: the paired-capture protocol is the load-bearing assumption, and it isn't documented in the text we have.\n\nWhat's genuinely new: the CL3AN dataset concept. Existing benchmarks like ISTD, WSRD, and AMBIENT6K deal with white or natural light; extending to explicit colored sources with an ambient reference is a recognized gap. The authors know the area—they built AMBIENT6K—and they candidly note its limitation: the direct and ambient lighting parameters are coupled, which ties the reference to exposure correction for white-aligned light. That's a good setup for the pitch. The RLN2 network is a fairly standard Retinex-plus-frequency architecture, so the contribution is really the benchmark plus a competitive method, not a new paradigm.\n\nThe soft spots are the ones you'd expect. First, the dataset construction: the text describes Figure 2(D) but gives no capture protocol. No statement that the colored-light and ambient passes share the same camera, lens, exposure, white balance, or tone mapping, and no alignment procedure. The ambient reference is only well-defined if the scene is static and the imaging pipeline is identical apart from the light. If the colored pass introduces specular highlights, interreflections, or saturation that the ambient pass doesn't reproduce, the target is not a clean ground truth. The paper's own concession about AMBIENT6K's coupling makes this more pressing, not less. Second, the experimental sections are absent from this draft—no tables, metrics, ablations, or dataset statistics. So the 'extensive evaluations' claim is uncheckable from what we have.\n\nThat said, none of this is an internal contradiction. The paper reads as a competent group's next step in a line they've already published (AMBIENT6K, WSRD), and they promise code and models. The flaw is missing documentation, not incoherent reasoning.\n\nThe paper is for anyone working on image restoration, color constancy, or relighting: if the dataset survives scrutiny, it's an enabling resource. It deserves a serious referee, but the referee should push hard on the capture protocol and ask for the raw alignment and calibration details before the benchmark can be trusted.\n\nRecommendation: send to peer review, with the expectation that the experimental and dataset-construction sections will be the focus. I'd want to see a revised version before endorsing it.","headline":"A promising dataset idea whose credibility hinges on a capture protocol the draft doesn't provide—worth refereeing, but the reviewer should demand the calibration details.","tokens_in":9538,"tokens_out":2978,"would_cite":true,"duration_ms":33523,"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":"The paper argues that ambient lighting normalization can be extended from single white light to multiple colored lights, and offers the CL3AN dataset and the RLN2 network as the means to do it.","keywords":["Ambient Lighting Normalization","colored light sources","multi-illuminant scenes","illumination-reflectance decomposition","Retinex","paired dataset","chromaticity-luminance guidance","image restoration"],"falsifier":"Run dense correspondence between the colored-light and ambient versions of the same CL3AN scene: if static parts of the scene show residual geometric motion, or if flat patches of known color differ beyond illumination, the paired ground truth is not valid. That failure would falsify the dataset and the comparisons built on it.","tokens_in":8653,"feed_emoji":"💡","tokens_out":7218,"duration_ms":85301,"temperature":0.7,"pith_summary":"This paper tries to establish that Ambient Lighting Normalization can handle multiple colored light sources, not just the single white or white-balanced lighting assumed by most prior work. It introduces CL3AN, a dataset of paired captures in which the same scenes are recorded under multiple colored lights and under uniform ambient light, and claims this is the first large-scale, high-resolution dataset built for that task. Alongside the dataset, it proposes RLN2, a learning framework that restores the ambient-lit image by explicitly guiding a Retinex-style separation of illumination from reflectance using chromaticity and luminance components. If these claims hold, CL3AN provides a benchmark that exposes the failures of existing methods—illumination inconsistencies, texture leakage, and color distortion—and RLN2 offers a way to remove colored lighting artifacts at competitive computational cost.","feed_headline":"Multi-colored lighting can now be normalized to ambient light","feed_subtitle":"A large paired dataset and a Retinex-style model restore uniformly lit scenes at competitive cost.","key_machinery":"The load-bearing objects are the CL3AN capture protocol and the RLN2 network. In CL3AN, each scene is photographed once under multiple colored (RGB) light sources and once under uniform ambient light; the direct-lighting setup removes the color consistency constraint of earlier datasets so that complex material-light interactions appear. RLN2 then learns the mapping between the two captures by explicit chromaticity (color) and luminance (brightness) component guidance, a Retinex-inspired instruction that forces the network to separate illumination from reflectance rather than memorize a global color transform. That explicit decomposition is what the paper says lets the model avoid the artifacts seen in existing methods.","core_discovery":"The paper's central claim is that colored, multi-source lighting can be normalized to an ambient-lit reference without sacrificing robustness or speed, provided the model is told how to separate what the lights do from what the surfaces look like. The evidence offered is CL3AN, a large-scale, high-resolution paired dataset in which direct lighting is produced by RGB lights and the ambient reference is acquired under a separate uniform lighting setup, deliberately dropping the color consistency constraint used by earlier datasets. On top of it, the paper presents RLN2, which uses explicit chromaticity-luminance component guidance, inspired by the Retinex model, to perform the illumination-reflectance decomposition needed for restoration. According to the paper, benchmarking shows that leading approaches produce artifacts because they cannot disentangle illumination from reflectance, while RLN2 handles non-homogeneous color lighting and material-specific reflectance variations with competitive computational cost.","pith_inferences":["If CL3AN's pairing is sound, it could become a shared testbed for neighboring problems such as color constancy and white balance under multiple illuminants, since it provides controlled color-shifted inputs with known ambient references.","A natural extension would be to record CL3AN-style pairs as video or under varying light directions, turning the static normalization task into a relighting benchmark; the paper does not attempt this.","The explicit chromaticity-luminance guidance suggests that smaller, non-generative models may close much of the gap with diffusion-based restoration on color-dominated degradations, which would be a testable hypothesis on other restoration tasks."],"forward_implications":["Ambient Lighting Normalization can be evaluated under multiple colored light sources, not just single white or white-aligned lighting, making the task closer to real indoor and event scenes.","Existing restoration models trained on single-light or white-domain data can be measured and shown to produce illumination inconsistencies, texture leakage, and color distortion on CL3AN.","RLN2 can serve as a preprocessing step for applications that need illumination-invariant inputs, such as neural image editing, so that downstream editing or recognition sees reflectances rather than colored shadows.","Because RLN2 stays computationally competitive, colored-light normalization can be applied in practical settings where diffusion-based restoration is too slow."],"supporting_citations":[{"why":"Defines the Ambient Lighting Normalization task and supplies the multi-source white-light dataset and baseline that CL3AN and RLN2 build on and compare against.","marker":"[66]"},{"why":"The ISTD dataset provides the single-light shadow-removal benchmark whose assumptions the paper argues are too simple.","marker":"[69]"},{"why":"The WSRD dataset supplies the high-resolution laboratory shadow benchmark used as a comparison point for controlled capture setups.","marker":"[65]"},{"why":"The Retinex model is the theoretical inspiration for the illumination-reflectance decomposition that RLN2 is built around.","marker":"[43]"},{"why":"Retinexformer serves as a state-of-the-art Retinex-based baseline for image restoration and low-light enhancement.","marker":"[9]"},{"why":"Restormer is a strong transformer-based restoration baseline against which RLN2's performance and computational cost are measured.","marker":"[84]"},{"why":"SFNet provides a frequency-domain baseline that represents an alternative approach to feature refinement in restoration.","marker":"[12]"},{"why":"MambaIR supplies a state-space-model baseline representing the latest class of global-range restoration architectures.","marker":"[23]"}],"fun_headline_variants":["Normalizing colored lighting to ambient with paired data","Retinex-inspired model normalizes colored lighting to ambient","First paired dataset for ambient normalization under colored lights","CL3AN dataset and RLN2 model fix colored light artifacts","From RGB party lights to uniform glow: a dataset and model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim depends on the CL3AN paired captures: the colored-light and ambient-lit images of each scene must be pixel-aligned, and the ambient image must be a correct, lighting-independent ground truth for the same scene, yet the provided text does not show the capture, alignment, and post-processing details that would verify this.","fun_headline_variants_meta":{"raw":{"variants":["Normalizing colored lighting to ambient with paired data","Retinex-inspired model normalizes colored lighting to ambient","First paired dataset for ambient normalization under colored lights","CL3AN dataset and RLN2 model fix colored light artifacts","From RGB party lights to uniform glow: a dataset and model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001196,"raw_usage":{"total_tokens":4925,"prompt_tokens":931,"completion_tokens":3994,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":3914}},"tokens_in":547,"tokens_out":3994,"duration_ms":31914,"temperature":1.0,"reasoning_tokens":3914,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T05:06:07.763213+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run dense correspondence between the colored-light and ambient versions of the same CL3AN scene: if static parts of the scene show residual geometric motion, or if flat patches of known color differ beyond illumination, the paired ground truth is not valid. That failure would falsify the dataset and the comparisons built on it.","supporting_citations":[{"cited_title":"Stacked condi- tional generative adversarial networks for jointly learning shadow detection and shadow removal","cited_arxiv_id":null,"evidence_quote":"The ISTD dataset provides the single-light shadow-removal benchmark whose assumptions the paper argues are too simple."},{"cited_title":"Wsrd: A novel benchmark for high resolution image shadow removal","cited_arxiv_id":null,"evidence_quote":"The WSRD dataset supplies the high-resolution laboratory shadow benchmark used as a comparison point for controlled capture setups."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The Retinex model is the theoretical inspiration for the illumination-reflectance decomposition that RLN2 is built around."},{"cited_title":"Restormer: Efficient transformer for high-resolution image restoration","cited_arxiv_id":null,"evidence_quote":"Restormer is a strong transformer-based restoration baseline against which RLN2's performance and computational cost are measured."},{"cited_title":"Mambair: A simple baseline for im- age restoration with state-space model","cited_arxiv_id":null,"evidence_quote":"MambaIR supplies a state-space-model baseline representing the latest class of global-range restoration architectures."}],"review_version":1}