{"id":"437b3dd0-a1a0-4494-8d2d-5a435b19ccce","arxiv_id":"2412.08200","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"GN-FR uses a generalizable NeRF with flare-mask-guided attention and loss to render flare-free novel views from images corrupted by lens flare, and introduces a real multi-view flare dataset.","lead":"A team from IIT Madras built GN-FR, a neural rendering system that removes lens flare from multi-view photos and can generate new views of a scene. It is the first method to treat flare removal as a multi-view problem instead of fixing each image alone, and it reports large quality gains over prior single-image pipelines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Masking loss zeroes supervision exactly where flare removal matters; reported whole-image PSNR may hide failure to reconstruct flare regions.","rationale":"The reader's weakest assumption correctly identifies the coverage condition in Sec. 4.3 as load-bearing, and I partially agree. The masking loss zeroes out the loss at every target pixel where M=1, so those pixels receive no direct supervision. If the source views contain a flare-free observation of the same 3D point, the prediction can in principle be correct through the same function learned on M=0 pixels, because the target image is not an input to the renderer. But if no source view is flare-free at that projection, the Point Sampler masks all attention contributions and the prediction is unconstrained. The paper does not measure how often this happens on real scenes, nor does it report quantitative results restricted to flare regions. Table 1's whole-image PSNR can be dominated by the majority flare-free pixels, so the 26.18 dB figure does not establish flare-region reconstruction quality. The proposed check—computing PSNR separately in flare regions and measuring clean-source coverage—would settle whether the central claim actually holds. Given the missing evidence, the conditional verdict is appropriate; no verdict change is needed.","tokens_in":8352,"tokens_out":6393,"duration_ms":74646,"concrete_test":"On the synthetic IBRNet benchmark, where clean ground truth is available, compute PSNR/SSIM separately for pixels inside the flare-occupancy mask (M=1) versus outside, and compare the flare-region PSNR against the corrupted input's flare-region PSNR. Also compute, for each target flare pixel, whether at least one sampled source view has M=0 at its projected location, and report the fraction of target flare pixels with zero clean-source coverage. If flare-region PSNR is within a few dB of the corrupted input, or a nontrivial fraction of flare pixels has no clean source view, the masking loss does not actually remove flare and the central claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. 4.3 defines Lunsup = ||Pred ⊙ (1−M) − Target ⊙ (1−M)||, so the loss is identically zero at every target pixel where M=1—precisely the pixels the method claims to restore. The network receives no gradient telling it what clean color should be rendered under a target flare. If at least one sampled source view is flare-free at the corresponding 3D point, the prediction can in principle inherit clean information through the same function learned on M=0 pixels, since the target image is not an input to the renderer. But if no sampled source view is flare-free there, the Point Sampler (Eq. 3) multiplies all attention values for that point by zero; after softmax, attention redistributes over unrelated epipolar points and the rendered color becomes uncontrolled. The paper never measures this coverage condition on its real dataset, and Table 1 reports whole-image PSNR on synthetically flared IBRNet scenes, which can be dominated by the large flare-free majority. Thus the central claim of 'simultaneously eliminating flare artifacts' is not quantitatively supported for the pixels that actually contain flare.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GN-FR, a generalizable neural rendering framework built on GNT for novel view synthesis from flare-corrupted multi-view images, with the goal of jointly rendering novel views and removing lens flare in an unsupervised manner. The method consists of a Flare-occupancy Mask Generation network (FMG), a View Sampler, a Point Sampler, and a masking loss (Sec. 4.3, Eq. (4)) that supervises only flare-free target pixels. The authors also introduce a real 3D flare dataset of 17 scenes with 782 images, 80 real flare patterns, and annotated flare-occupancy masks. Quantitative evaluation is reported on synthetically flared IBRNet scenes (Table 1), where GN-FR reaches 26.18 dB PSNR versus 22.44 dB for vanilla GNT, and qualitative results are shown on synthetic and captured real scenes.","tokens_in":8545,"tokens_out":5264,"duration_ms":57020,"significance":"If the claims are substantiated, GN-FR would be the first multi-view, generalizable flare-removal method inside a NeRF-style framework, and the problem framing is timely and practical: flare is view-dependent, so neighboring views are a natural source of information. The idea of a masking loss to avoid paired flare/flare-free data is clever, and the authors deserve credit for capturing a real multi-view flare dataset, training a PSPNet-based mask generator with a 5:1 class-weight ratio, and reporting quantitative mask metrics (0.81 mIoU / 0.94 mAcc). The proposed View Sampler and Point Sampler are intuitive and the comparison against Flare7K++ pre-processing pipelines is relevant. However, the central claim is currently supported mainly by a single synthetic quantitative table with no error bars, no code or dataset release, no flare-region-specific metrics, and only qualitative real-scene validation. As presented, the evidence is not yet sufficient to establish the claimed generalization and simultaneous flare-removal performance.","major_comments":[{"comment":"The masking loss is identically zero at every target pixel where M=1, which are precisely the pixels that the method claims to restore. The paper itself acknowledges the required condition ('This presupposes that at least one of the source views remains unaffected by flare in the target region'), but it never measures this coverage condition. The authors should report, on both the synthetic and real datasets, the fraction of target flare pixels for which at least one sampled source view is flare-free, and should report PSNR/SSIM/LPIPS restricted to the flare-occupied regions (M=1). Without such metrics, Table 1's whole-image PSNR is dominated by the large flare-free majority and does not demonstrate that flare artifacts are actually removed.","section":"Sec. 4.3, Eq. (4)"},{"comment":"The quantitative evaluation is under-specified. The manuscript does not state whether the synthetically imposed flares on IBRNet scenes use the same 80 flare patterns that trained the FMG, whether the View Sampler's 10% flare-occupancy target exclusion (Sec. 4.2.1) is applied at evaluation time, how many scenes and patterns are used, or how many random trials are averaged. Error bars or per-scene variance are absent. If the test flares come from the same distribution used to train FMG, the result is an in-distribution evaluation and the 'cross-scene generalization' claim requires testing on held-out flare patterns or a separate synthetic protocol.","section":"Sec. 5.2, Table 1"},{"comment":"The real-scene validation is qualitative, and the model is fine-tuned for 50k iterations before testing on real scenes, but the manuscript does not state which real scenes are used for fine-tuning or whether they overlap with the evaluated scenes shown in Figs. 5-7. If fine-tuning is performed on the same scenes that are later evaluated, the results do not support a generalization claim. The authors should fine-tune on a separated subset of real scenes and evaluate on held-out real scenes, or clearly report that the qualitative real results are obtained without fine-tuning on the evaluated scenes.","section":"Sec. 5.1.2 and Fig. 7"},{"comment":"Reproducibility is currently insufficient for the central claim. No code, trained models, or dataset are released, and the training/inference protocol for the synthetic evaluation is incomplete (e.g., exact IBRNet scene split, number of source views at test time, mask thresholding details, and the relationship between the UNet feature encoder mentioned in Sec. 4 and the PSPNet-based FMG). The authors should release code and data, or at minimum provide a detailed protocol so that Table 1 can be independently reproduced.","section":"Sec. 5.1.1 and Sec. 5.1.2"}],"minor_comments":[{"comment":"Equation (1) has mismatched parentheses: the View-Transformer call is missing a closing parenthesis. Also, the notation for F_i(z, theta) could be clarified.","section":"Sec. 3, Eq. (1)"},{"comment":"In Eq. (3), A' = A * (1 - M), the dimensions of A and M and the per-source-view indexing are not defined. It should be stated explicitly that M is the flare occupancy mask of the corresponding source view and that the multiplication is element-wise.","section":"Sec. 4.2.2, Eq. (3)"},{"comment":"The loss is written as a norm without specifying whether it is the squared L2 norm; the text later calls it MSE loss. The equation should be updated to match the actual loss, e.g., ||Pred - Target||^2 with the mask applied.","section":"Sec. 4.3, Eq. (4)"},{"comment":"There are several typographical issues in the related-work section, including 'Wuet al.' (missing space), 'kotp et al.' (capitalization), and 'effected' should be 'affected.'","section":"Sec. 2.1"},{"comment":"The caption states 'on par with annotated masks,' which should be 'on a par with annotated masks.' Also, the abbreviation SOTA is used without definition, and the figure would benefit from quantitative ablation numbers rather than qualitative images alone.","section":"Sec. 5.4, Fig. 8"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses an interesting and timely problem, and the multi-view framing is a genuine contribution. However, the current evaluation does not yet substantiate the core claim of simultaneous novel-view synthesis and flare removal: the masking loss removes supervision exactly where flare removal matters, the only quantitative table is synthetic and lacks error bars and flare-region metrics, and the real-scene fine-tuning protocol is unclear. These issues are fixable with additional experiments and detailed reporting, so I recommend major revision rather than rejection. I would also encourage the editor to weigh the absence of code and dataset release when judging reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's the one-sentence take: the paper has a genuinely new problem—it treats flare as a view-dependent artifact and uses a generalizable NeRF to remove it from sparse views—but the evidence it offers does not yet support the strong clean-up claim, because the masking loss stops supervising exactly at the flare pixels.\n\nWhat is actually new: the multi-view formulation for flare removal is a real departure from the single-image Flare7K/Uformer line. The FMG/VS/PS modules are simple modifications of GNT, not deep new machinery, but they are sensible, and the ablations suggest VS and PS help. The dataset (17 real flare scenes, 782 images, 80 annotated flare patterns) is potentially the most useful contribution; if released, it could anchor future work. The FMG mask model reports 0.81 mIoU / 0.94 mAcc, which is evidence the mask cue is learnable. Citation pattern seems fine; the important prior works are there.\n\nWhere it gets soft. The masking loss in Eq. 4 is zero wherever M=1, so no gradient ever tells the network what clean color to render under a target flare. The paper states the required premise in Sec. 4.3—at least one source view is flare-free in that region—but never measures whether it holds. The stress-test is right that if all sampled source views have flare at the 3D point, the attention mask zeros everything and softmax redistributes to unrelated epipolar points. Whole-image PSNR on synthetically flared IBRNet scenes cannot reveal this because flare is a small fraction of pixels; a flare-region-only metric or a coverage analysis is needed. The real-scene \"finetuned for 50k iterations\" is written ambiguously and may include the same scenes shown in Figs. 6–7; if so those qualitative results are not generalization evidence. The synthetic test also uses the same flare-pattern distribution used to train FMG, which makes the favorable numbers less surprising. No code or dataset release is mentioned, no error bars, and the baseline set is thin (vanilla NeRF, vanilla GNT, Flare7K++ pipelines). None of these are fatal to the idea; they are fatal to the current evidence level.\n\nWho it is for: anyone working on scene-level image restoration, robust novel view synthesis, or adverse-condition rendering. It deserves a serious referee, but a skeptical one who insists on per-mask metrics and a clean fine-tuning protocol. My own verdict: conditional, leaning toward revise-and-resubmit if the authors can address the masking-coverage concern.\n\nWould I bring it to reading group? Maybe, mainly to debate the masking-loss coverage condition. I would not cite it yet until code or data appear.\n\nRecommendation: send to peer review; demand the masking-coverage analysis and clarify the fine-tuning protocol before acceptance.","headline":"Genuinely new problem formulation, but the masking loss zeroes supervision at exactly the flare pixels, and the current evaluation does not support the claimed cleanup.","tokens_in":9095,"tokens_out":3993,"would_cite":false,"duration_ms":43981,"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":"This paper proposes GN-FR, a generalizable neural radiance field framework that renders flare-free novel views from sparse, flare-corrupted multi-view inputs and reports 26.18 dB PSNR on synthetic flared IBRNet scenes.","keywords":["lens flare removal","neural radiance fields","generalizable NeRF","multi-view image-based rendering","flare occupancy mask","novel view synthesis","unsupervised learning","GNT"],"falsifier":"For each target view in the released multi-view flare dataset, compute the pixelwise intersection of the selected source views' flare masks; any pixel that is flared in every source view violates the coverage assumption and is rendered with no data-driven supervision, so inspecting those pixels would directly test whether the central premise holds.","tokens_in":8163,"feed_emoji":"📷","tokens_out":4049,"duration_ms":39472,"temperature":0.7,"pith_summary":"This paper argues that lens flare, though hard to remove from a single image, can be removed by treating it as a multi-view problem: because flare changes with viewpoint, the scene content hidden behind flare in one image is usually visible in neighboring images. The authors build GN-FR, a generalizable neural radiance field built on the Generalizable NeRF Transformer (GNT) architecture, which renders novel views from sparse flare-corrupted inputs while simultaneously removing the flare. To train without paired flare-free ground truth, they supervise only pixels classified as flare-free, using a learned flare-occupancy mask, and they contribute a real 3D multi-view flare dataset with annotated masks. On synthetically flared IBRNet scenes, GN-FR reaches 26.18 dB PSNR, about 3.7 dB above its non-flare-aware generalizable baseline.","feed_headline":"Multi-view trick removes lens flare from new viewpoints","feed_subtitle":"GN-FR uses neighboring flare-free pixels to clean halos and streaks while rendering novel views","key_machinery":"The machinery is the transformer-based generalizable NeRF (GNT) plus three flare-aware modifications. Flare-occupancy masks from a trained PSPNet are multiplied into the view transformer's attention ($A' = A*(1-M)$) and are used to select source views and to mask the loss. The load-bearing identity is the masking loss, which forces the target view to render only from pixels that are flare-free in the source views.","core_discovery":"The central claim is that the view-dependent nature of flare can be exploited to eliminate flare artifacts during novel view synthesis without needing flare-free reference images. The framework consists of three modules inserted into GNT: a Flare-occupancy Mask Generator (FMG) that labels flare regions, a View Sampler (VS) that selects source views with the least flare, and a Point Sampler (PS) that zeros attention weights for sampled points falling in flared regions. A masking loss $L_{\\mathrm{unsup}} = \\|\\mathrm{Pred} \\odot (1-M) - \\mathrm{Target} \\odot (1-M)\\|$ restricts training supervision to flare-free pixels, so the network learns to fill flared areas from nearby clean observations. The paper reports 26.18 dB PSNR on flared IBRNet scenes, exceeding the best baseline by about 4 dB, and demonstrates qualitative removal on real scenes.","pith_inferences":["The same masked-attention mechanism could be adapted to other view-dependent disturbances such as reflections, glare, or partial occluders, since the only requirement is that a mask identifies the artifact and that it disappears from some viewpoints.","The coverage premise could be checked directly on the released dataset: per pixel, count how many source views are flare-free; any target pixel whose flared mask is 1 in every source view receives no supervision and its rendered value is unconstrained.","If the premise fails for large, coherent halos, an explicit regularization, for example a perceptual prior or a physics-based flare model, would be needed for those regions."],"forward_implications":["A single trained model can take several flare-corrupted views of an unseen scene and output flare-free novel views, without per-scene optimization or paired clean captures.","The masking-loss principle extends the NeRF idea of novel-view synthesis to novel-region synthesis: image areas corrupted in the target can be treated like unseen regions and filled from other views.","On quantitative evaluation, the method outperforms the alternative of first removing flare with Flare7K++ and then rendering with NeRF or GNT by roughly 4 dB PSNR.","The released dataset of 17 real scenes and 80 real flare patterns with masks provides a multi-view benchmark for flare removal research."],"supporting_citations":[{"why":"Supplies the generalizable transformer NeRF architecture (GNT) that GN-FR modifies.","marker":"[13]"},{"why":"IBRNet dataset provides the multi-view training scenes on which synthetic flare is imposed for quantitative evaluation.","marker":"[16]"},{"why":"Flare7K++ is the state-of-the-art single-image flare remover used as the pre-processing baseline in comparisons.","marker":"[4]"},{"why":"PSPNet is the semantic segmentation architecture trained as the FMG flare-occupancy mask generator.","marker":"[22]"},{"why":"the 24K Flickr dataset is used to synthesize diverse flare-corrupted training images with random affine transformations.","marker":"[21]"},{"why":"vanilla NeRF serves as the scene-specific baseline that cannot generalize across scenes.","marker":"[10]"}],"fun_headline_variants":["NeRF model erases lens flare using neighboring views","Flare-free rendering from multi-view NeRF","Leveraging view-dependence to clean lens flare","GN-FR: NeRF that removes flare without clean references","Multi-view NeRF sweeps away lens flare artifacts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training loss supervises only flare-free pixels, so the method assumes every flared region in the target view is visible and flare-free in at least one sampled source view.","fun_headline_variants_meta":{"raw":{"variants":["NeRF model erases lens flare using neighboring views","Flare-free rendering from multi-view NeRF","Leveraging view-dependence to clean lens flare","GN-FR: NeRF that removes flare without clean references","Multi-view NeRF sweeps away lens flare artifacts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000514,"raw_usage":{"total_tokens":2523,"prompt_tokens":1001,"completion_tokens":1522,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":617,"completion_tokens_details":{"reasoning_tokens":1445}},"tokens_in":617,"tokens_out":1522,"duration_ms":12043,"temperature":1.0,"reasoning_tokens":1445,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:05:31.811469+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For each target view in the released multi-view flare dataset, compute the pixelwise intersection of the selected source views' flare masks; any pixel that is flared in every source view violates the coverage assumption and is rendered with no data-driven supervision, so inspecting those pixels would directly test whether the central premise holds.","supporting_citations":[{"cited_title":"Is attention all that neRF needs? In The Eleventh International Conference on Learning Representations , 2023","cited_arxiv_id":null,"evidence_quote":"Supplies the generalizable transformer NeRF architecture (GNT) that GN-FR modifies."},{"cited_title":"Ibrnet: Learning multi-view image-based rendering","cited_arxiv_id":null,"evidence_quote":"IBRNet dataset provides the multi-view training scenes on which synthetic flare is imposed for quantitative evaluation."},{"cited_title":"Flare7K++: Mixing Synthetic and Real Datasets for Nighttime Flare Removal and Beyond","cited_arxiv_id":"2306.04236","evidence_quote":"Flare7K++ is the state-of-the-art single-image flare remover used as the pre-processing baseline in comparisons."},{"cited_title":"Srinivasan, Matthew Tancik, Jonathan T","cited_arxiv_id":null,"evidence_quote":"vanilla NeRF serves as the scene-specific baseline that cannot generalize across scenes."}],"review_version":1}