{"id":"4848d575-b1f5-4903-961e-15d906d9e750","arxiv_id":"2507.17489","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DFDNet applies learnable dynamic frequency-domain filtering plus local contrastive guidance to remove lens flare, reporting state-of-the-art scores on Flare7K++ and real-world night images.","lead":"A new neural network, DFDNet, removes lens flare from night photos by filtering image features in the frequency domain and using contrastive learning to preserve details near light sources. On the Flare7K++ benchmark it reports higher PSNR and structural similarity scores than previous flare-removal methods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 4 ablation baseline is numerically identical to published Flare7k++ results, so the claimed gains from GDFG/LDGM/frequency loss are not established; the central module-contribution claim is unsupported.","rationale":"The reader's weakest assumption correctly identifies the ablation baseline issue, and I agree that it is the most load-bearing concern. The central scientific contribution of DFDNet is the claim that the GDFG and LDGM modules, together with the frequency loss, cause the performance improvement; the ablation is the only evidence for this. If the baseline row is simply the published Flare7k++ result rather than a model retrained under identical conditions, then the ablation is not internally controlled and cannot support any attribution. The exact match to three decimals makes it very unlikely that the baseline was retrained with a different seed or setup. This is not an accusation of misconduct; it is a request for evidence. The paper does not report error bars, seeds, or whether checkpoints are the official ones, so the possibility of an uncontrolled baseline is real. A concrete retraining experiment would settle it. I also note that the paper's statement 'our method achieves optimal performance across all the metrics' is contradicted by its own Table 1 (FPro has higher real S-PSNR) and Table 2 (FPro has higher MUSIQ). This overclaim is secondary; it can be corrected by rewording, whereas the ablation problem requires new experiments. Therefore, the appropriate verdict remains conditional: the method may work, but the evidence for the modules' contributions must be re-established with a controlled ablation and code release. No verdict change from the reader's conditional acceptance is needed.","tokens_in":16444,"tokens_out":9475,"duration_ms":89679,"concrete_test":"Request the official training code and checkpoints from the authors; retrain the Uformer backbone (same architecture as DFDNet minus GDFG/LDGM/frequency loss) using the exact same data pipeline, augmentation, loss, optimizer, and iteration schedule as DFDNet, and evaluate on the Flare7K++ real test set. If the retrained baseline matches Table 4's first row within numerical precision, verify whether it is the official Flare7k++ checkpoint; if so, the ablation is not controlled. In addition, run the full DFDNet and the retrained baseline with at least 3 random seeds and report mean ± std; check whether the reported GDFG gain (0.107 dB) and frequency-loss gain (0.211 dB) exceed the seed variance. Also re-read Table 1 to confirm FPro's real S-PSNR (23.405) versus DFDNet's 23.300; if correct, revise the 'all metrics' claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that DFDNet's modules (GDFG, LDGM, frequency loss) drive the improvement rests entirely on the ablation in Table 4. The '× × ×' baseline row reports PSNR 27.633, SSIM 0.894, LPIPS 0.0428, G-PSNR 23.949, S-PSNR 22.603 on the Flare7K++ real test set. These are exactly the numbers listed for Flare7k++ (TPAMI'24) in Table 1, to three decimal places. A Uformer baseline retrained under the same pipeline and augmentation would not be expected to coincide with a published method's metrics to this precision unless it is literally the same checkpoint or an exact copy of the published result. If the baseline is not a controlled retraining, then the incremental gains (+0.107 PSNR for GDFG, +0.019 for LDGM, +0.211 for frequency loss, +0.397 combined) may be due to differences in training protocol, data split, augmentation, or evaluation details rather than the modules. This leaves the paper's core contribution unverified. Note also that the paper's claim of 'optimal performance across all metrics' is contradicted by its own Table 1 (FPro achieves higher real S-PSNR: 23.405 vs 23.300) and Table 2 (FPro higher MUSIQ: 64.901 vs 64.702), indicating the SOTA claim is at least overstated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes DFDNet, a Uformer-style network for nighttime lens-flare removal, with two main components: a global dynamic frequency-domain guidance (GDFG) module that applies learnable, input-dependent spectral filters, and a local detail guidance module (LDGM) based on contrastive learning. The training objective combines L1/MSE perceptual losses, a frequency-domain L1 loss on magnitude and phase, and the LDGM contrastive loss. Experiments are reported on Flare7K++ real and synthetic test sets, on unpaired real-world benchmarks, and in ablations over the modules, the number of frequency filter bases N, the LDGM temperature tau, and the loss weights alpha and lambda. The paper claims state-of-the-art results on most paired metrics and on the unpaired no-reference metrics.","tokens_in":16757,"tokens_out":4285,"duration_ms":48368,"significance":"If the central claims are established, the paper would make a useful contribution to flare removal: the frequency-domain analysis in Figures 1 and 2 is a plausible motivation, the GDFG module is clearly specified with an algorithmic description, the code is promised publicly, and the reported numbers on Flare7K++ show consistent gains on PSNR/SSIM/LPIPS/G-PSNR over the strongest prior methods. The frequency loss in Eq. (10) is a standard supervised objective and does not introduce circularity. The main significance rests on the ablations in Tables 4-7, which currently do not establish the contribution of the individual modules because the baseline is not demonstrably a controlled retraining.","major_comments":[{"comment":"The ablation baseline row ('x x x') reports PSNR 27.633, SSIM 0.894, LPIPS 0.0428, G-PSNR 23.949, and S-PSNR 22.603 on the Flare7K++ real test set. These are exactly the values listed for the published Flare7K++ method in Table 1, to three decimal places. If the baseline is simply the published Flare7K++ result rather than a Uformer-based model retrained under the same data pipeline, augmentation, and training protocol as the full DFDNet, then the incremental gains attributed to GDFG, LDGM, and the frequency loss are not internally controlled. The paper should retrain the baseline under identical conditions and report the numbers; otherwise the central module-contribution claim is unsupported.","section":"Table 4"},{"comment":"As written, Eq. (7) is not the standard InfoNCE contrastive loss. The denominator is printed as exp(v·v+/tau + sum_n exp(v·v_n^-/tau)), which makes the ratio equal to exp(-sum_n exp(v·v_n^-/tau)), so increasing similarity to negatives would not behave as described in the text. The intended expression is presumably exp(v·v+/tau) / [exp(v·v+/tau) + sum_n exp(v·v_n^-/tau)]. The correct formula should be given, and the implementation should be checked against it, because the LDGM contribution is a central claim.","section":"Eq. (7)"},{"comment":"The hyperparameters N, tau, alpha, and lambda are each selected by comparing G-PSNR and S-PSNR computed on the Flare7K++ test set. There is no validation split or cross-validation, and no error bars or repeated-run variance are reported. Selecting hyperparameters directly on the test set risks overfitting the reported metrics and makes the claimed margins difficult to interpret. The authors should either use a held-out validation set for model selection or report the uncertainty and demonstrate that the conclusions are stable.","section":"Sec. 4.3.1, Tables 5-7"},{"comment":"The statement that 'our method achieves optimal performance across all the metrics' is contradicted by the paper's own tables. In Table 1, FPro achieves a higher real S-PSNR (23.405 vs 23.300), and in Table 2, FPro achieves a higher MUSIQ (64.901 vs 64.702). The claim should be revised to a precise statement of which metrics are best and where DFDNet is not first, and the discussion should address those cases.","section":"Sec. 4.2.1, Table 1"}],"minor_comments":[{"comment":"The 'Ours dagger' row, which uses the Flare7K dataset as the training set, is not discussed in the text. If it is intended as a comparison of training-data influence, it should be described; otherwise it may confuse the reader.","section":"Table 1"},{"comment":"The frequency loss uses L1 on amplitude and phase after the DFT. It would be helpful to state explicitly whether the phase is wrapped or unwrapped, since L1 on wrapped phase can behave discontinuously.","section":"Sec. 3.4 / Eq. (10)"},{"comment":"The notation for the dynamic weight sum in Algorithm 1 starts at i=1 in the displayed 'W = sum_{i=1}^N T_i Phi_i' but the previous text uses a similar expression with i=0; the indexing should be made consistent.","section":"Algorithm 1"},{"comment":"Several minor typographical issues appear, including 'Conact' for concatenation near Eq. (2) and inconsistent capitalization in Section titles such as 'Frequent Analysis' in Section 2.2. A careful proofread would improve readability.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The load-bearing weakness is the Table 4 ablation: the baseline numbers are identical to the published Flare7K++ row, so the module-level attribution is not currently established. This is fixable by a controlled retraining, which is why I recommend major revision rather than rejection. I would also ask the editor to ensure the authors provide the corrected Eq. (7) and clarify the validation procedure for hyperparameter selection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know about this paper is that its central ablation is not controlled. The baseline row in Table 4 is numerically identical, to three decimals, to the published Flare7k++ result in Table 1 (27.633 PSNR, 0.894 SSIM, 23.949 G-PSNR, 22.603 S-PSNR). That strongly suggests the authors used the published checkpoint or numbers rather than retraining a Uformer under their own pipeline. So the reported +0.397 PSNR from adding GDFG, LDGM, and the frequency loss cannot be attributed to those modules. The module-contribution claim is the core of the paper, and it is unsupported.\n\nWhat is actually new: GDFG is a dynamic global filter where an MLP with softmax combines N learnable filter bases per channel, and LDGM is a contrastive loss on local patches around the light source. That is an incremental but reasonable extension of existing frequency-domain filters (FPro, MFSNet, SFIR). The frequency-domain intuition about flares showing stronger artifacts than in the spatial domain is well-motivated, and the architecture is clearly described. The paper also evaluates on multiple benchmarks, which is good.\n\nSoft spots: besides the broken ablation, hyperparameters N, tau, alpha, lambda are all selected on the test set, and no error bars are reported. The code is only promised. The claim of \"optimal performance across all metrics\" is false: FPro has higher real S-PSNR (23.405 vs 23.300) and higher MUSIQ (64.901 vs 64.702) in their own tables. That overstatement should have been caught.\n\nWhat holds up: the architecture and losses are coherent, and the qualitative figures show plausible improvements. The idea of jointly using frequency-domain global guidance and local contrastive learning is sensible. The limitation section is honest about small light source detail loss.\n\nBottom line: this is a competent engineering paper with a flawed validation protocol. The task is real, the method is plausible, but the evidence does not yet establish the contributions. A careful referee could ask for a controlled retraining of the baseline, error bars or a significance test, and a corrected SOTA statement. That is a major revision, not a desk reject.\n\nRecommendation: send it to peer review. The flaws are fixable and the method deserves a fair test with proper controls.","headline":"Solid engineering but the ablation baseline is copied from published Flare7k++ numbers, so the module contributions are unproven; the SOTA claim is also overstated.","tokens_in":17316,"tokens_out":2500,"would_cite":false,"duration_ms":25617,"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":"Flare artifacts are easiest to separate in the frequency domain, and a network built on that idea beats all compared deflare methods.","keywords":["flare removal","frequency domain","contrastive learning","image restoration","nighttime photography","dynamic frequency filter","Uformer architecture","Fourier transform"],"falsifier":"Retrain the Table 4 baseline (Uformer without GDFG, LDGM, or the frequency loss) using the paper's own code and pipeline, and compare its Flare7K++ real-test PSNR with the reported 27.633 dB; if the retrained baseline comes out materially above that value or if DFDNet's margin over it shrinks without the modules, the attribution of the improvements to GDFG and LDGM fails. An equally direct check is to feed the predicted flare maps from GDFG against the annotated masks in Flare7K++: if the maps do not highlight the marked glare and streak regions better than the backbone's output, the decoupling claim is unsupported.","tokens_in":16247,"feed_emoji":"✨","tokens_out":8156,"duration_ms":77523,"temperature":0.7,"pith_summary":"DFDNet claims that the hardest part of removing nighttime lens flares is not the visible streaks but their Fourier spectrum: flare artifacts widen the spectrum's energy and paint radial stripes onto it, making them easier to disentangle from real content in the frequency domain than in pixel space. To exploit that, the network inserts a global dynamic frequency-domain guidance (GDFG) module into every stage of a U-shaped transformer; the module multiplies each channel's Fourier transform by weights that are generated from the image's own channel statistics, so the filtering adapts to the specific flare. A local detail guidance module then uses contrastive learning to keep the restoration from erasing small structures around the light source. The paper reports state-of-the-art results on the Flare7K++ real and synthetic test sets, on a flare-corrupted real-world set, and on a consumer-electronics test set, and shows that deflaring improves downstream detection and segmentation.","feed_headline":"Frequency-guided deflare hits 28.03 dB PSNR, beating all prior methods","feed_subtitle":"Dynamic Fourier-domain filtering and contrastive learning remove large nighttime flares while preserving light sources.","key_machinery":"The load-bearing object is the global dynamic frequency-domain guidance module (GDFG). It applies a discrete Fourier transform to the feature map, multiplies the spectrum by a multichannel weight $W = \\sum_i T_i(X) \\Phi_i$, where $\\Phi_i$ are $N$ learnable complex masks and $T_i$ are softmax-normalized coefficients produced by global average pooling followed by an MLP, then applies the inverse transform and adds a residual connection. Because the weights depend on the input, the filter is content-adaptive rather than a fixed global filter. The second component is the local detail guidance module (LDGM), a contrastive loss on randomly cropped patches that treats restored-image patches as queries, co-located reference patches as positives, and the remaining patches as negatives. The paper also uses a frequency-domain loss $L_{FFT}$ that takes L1 distances between the amplitudes and phases of the restored and reference images' Fourier transforms.","core_discovery":"The central discovery the authors are trying to establish is that content and flare artifact information can be decoupled in the frequency domain by a learnable, input-adaptive filter, and that this decoupling transfers to real images. Concretely, the paper claims that large-scale flares add broadband energy and directional radial patterns to the mid- and high-frequency spectrum, that a per-channel dynamic Fourier mask computed from global average pooling and an MLP can suppress those components, and that a phase-and-amplitude L1 loss in the Fourier domain sharpens this separation. The local contrastive module is presented as the counterpart that realigns damaged patches near the light source with the reference image, fixing the detail loss that pure frequency filtering would cause. The combined system is claimed to outperform every compared method on all reported paired and unpaired benchmarks, with e.g. 28.030 dB PSNR on the Flare7K++ real test set versus 27.662 dB for the closest competitor.","pith_inferences":["The Table 4 ablation baseline is numerically identical to the published Flare7K++ row in Table 1 (PSNR 27.633, SSIM 0.894, G-PSNR 23.949, S-PSNR 22.603); if that row is the original model rather than a Uformer retrained under the paper's exact pipeline, then the reported per-module gains (e.g. +0.107 dB for GDFG) mix architectural benefit with training-protocol differences, so the isolated module ","The same 'dynamic Fourier mask from channel statistics' recipe is a general mechanism: any optical artifact with a directional spectral signature, such as anamorphic flare, diffraction spikes, or raindrop streaks, could be attacked with the same input-adaptive frequency weighting instead of the fixed filters used in existing restoration networks.","A cheap, testable extension is to take a pretrained deflare network that lacks frequency guidance and bolt on a GDFG module at its bottleneck; the paper's inference-time increase is only about 0.018 s, so the mechanism should be portable to other U-shaped restoration models."],"forward_implications":["If the reported comparisons are taken at face value, DFDNet is the current best deflare method on Flare7K++ real and synthetic tests, with the largest gains concentrated in G-PSNR and S-PSNR, the metrics that score flare-region fidelity.","The GDFG module alone improves the Uformer backbone's real-test PSNR from 27.633 to 27.740 dB, and adding the frequency loss raises it to 27.951 dB, so each proposed component contributes a measurable, separable improvement.","The combination of GDFG and LDGM yields a predicted flare map that covers annotated glare, streak, and light-source regions without misclassifying the light source itself, which the paper presents as evidence that global frequency guidance and local contrastive alignment are complementary.","Because flare removal improves object detection and semantic segmentation in the paper's downstream experiments, the restoration quality transfers to practical vision tasks rather than only to image-quality metrics."],"supporting_citations":[{"why":"Supplies the paired training pipeline, the real and synthetic test sets, and the G-PSNR/S-PSNR flare-region metrics on which the headline comparison rests.","marker":"[9]"},{"why":"Provides the original nighttime flare dataset and the flare/light-source pool that the method and its competitors train on.","marker":"[8]"},{"why":"The U-shaped transformer backbone that DFDNet modifies by inserting GDFG modules, and the architecture used for the Table 4 ablation baseline.","marker":"[39]"},{"why":"The closest prior competitor on the real Flare7K++ test set (27.662 dB PSNR), which the method must beat and does by 0.368 dB.","marker":"[24]"},{"why":"A recent frequency-prompt restoration method compared against in Tables 1 and 2, setting the frequency-domain baseline to which GDFG is contrasted.","marker":"[51]"},{"why":"The synthesis-based training approach for flare removal from which the field's paired-data pipeline descends, and a baseline in Table 1.","marker":"[41]"}],"fun_headline_variants":["Dynamic Fourier masks strip large night flare artifacts","Frequency decoupling removes large flares, new SOTA","Frequency-guided deflare hits 28.03 dB, new state of the art","Fourier-domain guidance outperforms all deflare methods","Adaptive frequency filtering eliminates large artifacts, beats SOTA"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Table 4 baseline row is a Uformer retrained from scratch under the same data pipeline, augmentation, and training protocol as the full DFDNet; if that row is instead the original published Flare7K++ model, the reported gains from GDFG and LDGM are not attributable to the modules.","fun_headline_variants_meta":{"raw":{"variants":["Dynamic Fourier masks strip large night flare artifacts","Frequency decoupling removes large flares, new SOTA","Frequency-guided deflare hits 28.03 dB, new state of the art","Fourier-domain guidance outperforms all deflare methods","Adaptive frequency filtering eliminates large artifacts, beats SOTA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0008,"raw_usage":{"total_tokens":3540,"prompt_tokens":988,"completion_tokens":2552,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":2468}},"tokens_in":604,"tokens_out":2552,"duration_ms":16425,"temperature":1.0,"reasoning_tokens":2468,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:47:21.726552+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the Table 4 baseline (Uformer without GDFG, LDGM, or the frequency loss) using the paper's own code and pipeline, and compare its Flare7K++ real-test PSNR with the reported 27.633 dB; if the retrained baseline comes out materially above that value or if DFDNet's margin over it shrinks without the modules, the attribution of the improvements to GDFG and LDGM fails. An equally direct check is to feed the predicted flare maps from GDFG against the annotated masks in Flare7K++: if the maps do not highlight the marked glare and streak regions better than the backbone's output, the decoupling claim is unsupported.","supporting_citations":[{"cited_title":"Flare7k ++: Mixing synthetic and real datasets for nighttime flare removal and beyond","cited_arxiv_id":null,"evidence_quote":"Supplies the paired training pipeline, the real and synthetic test sets, and the G-PSNR/S-PSNR flare-region metrics on which the headline comparison rests."},{"cited_title":"Flare7k: A phenomenological nighttime flare removal dataset","cited_arxiv_id":null,"evidence_quote":"Provides the original nighttime flare dataset and the flare/light-source pool that the method and its competitors train on."},{"cited_title":"Uformer: A general u-shaped transformer for image restora- tion","cited_arxiv_id":null,"evidence_quote":"The U-shaped transformer backbone that DFDNet modifies by inserting GDFG modules, and the architecture used for the Table 4 ablation baseline."},{"cited_title":"Flare-free vision: Empowering uformer with depth insights","cited_arxiv_id":null,"evidence_quote":"The closest prior competitor on the real Flare7K++ test set (27.662 dB PSNR), which the method must beat and does by 0.368 dB."},{"cited_title":"Seeing the unseen: A frequency prompt guided transformer for im- age restoration","cited_arxiv_id":null,"evidence_quote":"A recent frequency-prompt restoration method compared against in Tables 1 and 2, setting the frequency-domain baseline to which GDFG is contrasted."},{"cited_title":"How to train neural networks for flare re- moval","cited_arxiv_id":null,"evidence_quote":"The synthesis-based training approach for flare removal from which the field's paired-data pipeline descends, and a baseline in Table 1."}],"review_version":1}