{"id":"881315bd-e3e6-4c57-9d4c-e837c2473383","arxiv_id":"2504.14664","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A frequency-domain fusion module that injects estimated blur kernels into a Transformer improves blind image deblurring on in-domain and out-of-domain benchmarks.","lead":"This paper adds blur-kernel information into a frequency-domain Transformer network for image deblurring, using a new Frequency Integration Module. The authors report improved deblurring quality and generalization on standard benchmarks including GoPro, HIDE, and RealBlur.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 3 ablation baseline is 1.66 dB below the retrained FFTFormer; without an explanation, the 0.74 dB gain attributed to the kernel prior is not interpretable.","rationale":"The reader's weakest_assumption identifies the same load-bearing issue: the Table 3 no-prior baseline is unexplained and inconsistent with the retrained FFTFormer number in Sec. 4.2.1. This is the central empirical support for the kernel prior's effect, because Table 1 alone conflates the kernel prior with added parameters, FIM blocks, multi-scale integration, and a three-stage training schedule. The 1.66 dB gap between 32.45 and 34.11 must be resolved before the 0.74 dB gain can be attributed to FIM. If the protocols differ, the paper must disclose the training budgets; if they do not, the baseline is likely misreported. The paper's idea—introducing a domain-independent kernel prior through frequency-domain fusion—is plausible, and the out-of-domain improvements on HIDE and RealBlur-R are encouraging, so this is not grounds for rejection. However, the controlled evidence has an unresolved inconsistency, so the CONDITIONAL verdict stands. A code release or a clear statement of the exact training schedules would settle the issue.","tokens_in":11761,"tokens_out":10627,"duration_ms":85710,"concrete_test":"Retrain the FFTFormer backbone without FIM and without the kernel estimator under two protocols: (A) exactly the Table 3 protocol (same iterations, loss, batch size, resolution) and (B) the Sec. 4.2.1 protocol (original FFTFormer settings). Compare the GoPro PSNRs to 32.45 and 34.11 respectively. If protocol A reproduces 32.45 and protocol B reproduces 34.11, the gap is a training-budget difference, making Table 3 internally valid but Table 1 compute-imbalanced. If protocol A reproduces ~34.11, the Table 3 baseline is misreported and the claimed 0.74 dB kernel-prior gain is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's controlled evidence for the kernel prior is Table 3. Row 1 ('end-to-end training without introducing kernel prior') reports 32.45 dB on GoPro, and row 6 (full FIM) reports 33.19 dB, implying a 0.74 dB gain. But Sec. 4.2.1 says the authors retrained the same backbone (FFTFormer) and obtained a result 0.05 dB below the published 34.16, i.e., about 34.11 dB. The 1.66 dB gap between the no-prior baseline and the retrained backbone is never explained. The Table 3 caption only guarantees the same setting within that table, not that it matches the Sec. 4.2.1 protocol. If the no-prior baseline under the Table 3 protocol should actually be near 34.11, then the FIM improvement becomes 33.19 - 34.11 = -0.92 dB, contradicting the claim. If the protocol differs (e.g., fewer iterations), then Table 1's comparison of the full model (which has extra training stages) to FFTFormer is confounded by compute, and the 0.10 dB GoPro gain is especially fragile. Either way, the attribution of the improvement to the kernel prior is not cleanly supported. The value 32.45 also coincides exactly with MIMO-UNet+'s GoPro PSNR in Table 1 (SSIM differs), suggesting a possible transcription error.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a new architecture for blind image deblurring that combines an explicit blur kernel estimation module with a frequency-domain Transformer backbone. The kernel prior is fused into the network through a Frequency Integration Module (FIM) that performs a cross-attention-like operation in the Fourier domain. Training is conducted in three stages: pretraining the kernel estimator with a reblurring loss, training the backbone and FIM with the kernel estimator frozen, and joint fine-tuning with an additional reblurring loss. The authors evaluate on GoPro, HIDE, RealBlur-R, and RealBlur-J, reporting state-of-the-art results and claiming improved out-of-domain generalization. The central claim is that the explicit kernel prior improves deblurring performance and generalization.","tokens_in":12007,"tokens_out":6212,"duration_ms":50253,"significance":"If the reported results are reproducible, the method provides a modest improvement over the FFTFormer backbone on the in-domain GoPro benchmark (+0.10 dB) and larger gains on out-of-domain HIDE (+0.46 dB) and RealBlur-R (+1.13 dB). The idea of using a learned kernel prior for domain generalization is of interest to the image restoration community. The paper does not provide code or pre-trained models, and no machine-checked proofs or parameter-free derivations are involved. The main significance hinges on the ablation study in Table 3, which currently contains an unexplained baseline discrepancy that must be resolved before the improvement can be attributed to the kernel prior.","major_comments":[{"comment":"The first row of Table 3 is described as 'end-to-end training without introducing kernel prior' and reports 32.45 dB PSNR on GoPro, yet Section 4.2.1 states that the authors retrained the same FFTFormer backbone and obtained 34.11 dB PSNR. This 1.66 dB discrepancy is not explained. The Table 3 caption says all options are trained with the same setting, but the relationship between that setting and the Section 4.2.1 retraining protocol is not specified. If the baseline is not the backbone under the full model's training protocol, the 0.74 dB improvement attributed to FIM (row 6, 33.19 dB) is not interpretable. The value 32.45 dB also exactly matches MIMO-UNet+'s GoPro PSNR in Table 1, raising the possibility of a transcription error. Please clarify the training protocol and correct or explain the numbers.","section":"Table 3, row 1 vs Section 4.2.1"},{"comment":"The proposed Frequency Attention operation is not fully specified. F(Q) and F(K) are complex-valued, and their pointwise product is complex; the inverse FFT generally produces a complex output, but layer normalization and element-wise multiplication with V are typically defined for real tensors. The paper does not state whether the real part is taken, nor how the kernel feature B_in is projected to match the spatial and channel dimensions of the image feature Q. These implementation details are essential for reproducibility and for verifying that the operation behaves as intended.","section":"Section 3.2, Eq. (3)"},{"comment":"The non-blind ablation compares the proposed method with ground-truth kernels against a 'Baseline' (without kernel prior) on synthetic datasets. However, the synthetic dataset construction (extracting kernels from the original datasets and reblurring the corresponding sharp images) is described only briefly. The reported PSNR values for the same dataset names (e.g., GoPro Syn. 28.48 dB vs. GoPro test 32.45 dB in Table 3) are very different, and it is unclear whether these synthetic sets are derived from the training or test partitions. Without this clarification, the non-blind result does not provide a clean measure of the potential benefit of a more accurate kernel estimator.","section":"Section 4.3, Table 4"},{"comment":"In stage III, the reblurring loss L_KE is computed using the predicted deblurred image x_hat_n from the backbone rather than the ground truth x_n. Since x_hat_n is itself an output being trained, the kernel estimator is optimized against a target that depends on the current deblurring quality. The paper does not analyze the potential interaction or instability of this joint optimization, and no sensitivity analysis is provided for the weights lambda_1 and lambda_2 (both set to 0.1). This is a load-bearing point because the reblurring loss is central to the self-supervised kernel estimation that supports the generalization claim.","section":"Section 3.3, Eq. (7)"}],"minor_comments":[{"comment":"The heading 'Qualitative Comparison' is misleading because the section reports quantitative results (PSNR/SSIM); the subsequent Section 4.2.2 'Visual Comparison' is the qualitative part. Please rename the sections accordingly.","section":"Section 4.2.1"},{"comment":"The use of double quotes to indicate which options are enabled is unconventional and hard to read. Consider using checkmarks or a different formatting convention.","section":"Table 3"},{"comment":"The naming of synthetic datasets is inconsistent: the text says 'synthetic datasets are respectively based on GoPro, HIDE, RealBlur-R and RealBlur-J', while Table 4 uses 'GoPro (Syn.)', 'HIDE (Syn.)', 'RB-R (Syn.)', and 'RB-J (Syn.)'. Please standardize the terminology and clarify the relationship to the original test sets.","section":"Section 4.3"},{"comment":"The paper states that stage II 'follows the training settings of [12]' but does not report the number of training iterations for stage II. Since the three-stage training schedule is a key part of the method, this information should be included.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The discrepancy between Table 3 row 1 and the retrained FFTFormer result in Section 4.2.1 is the main concern, and the exact match of 32.45 dB with MIMO-UNet+ in Table 1 suggests a possible copy-paste error. It would be helpful if the editor could request the authors to release the exact configuration of the Table 3 baseline and, ideally, the code or detailed training logs to allow verification of the central ablation claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read on the deblurring paper: the Frequency Integration Module is a genuine new piece of engineering, and the out-of-domain numbers on HIDE and RealBlur-R are worth a look. But the ablation table has a baseline problem that makes the claimed contribution hard to interpret.\n\nWhat's actually new: FIM fuses kernel and image features in the frequency domain via an attention-like operation, inserted multiscale into the FFTFormer backbone. That's a reasonable derivative of the FFTFormer line, but the specific fusion formula is not in the kernel-based deblurring literature I know. The three-stage training schedule with reblurring loss for the kernel estimator is standard but thoughtfully combined.\n\nWhat it does well: The paper is clearly written, compares against a retrained FFTFormer (they admit they couldn't reproduce the published number exactly), and the visual comparisons are consistent with the quantitative gains on HIDE and RealBlur-R. The non-blind oracle experiment in Table 4 is a nice check that a better kernel helps.\n\nThe main problem: Table 3 row 1—the no-kernel-prior baseline—reports 32.45 dB on GoPro, but Sec. 4.2.1 says the same backbone (FFTFormer) retrained with the paper's settings gives 34.11 dB. That's a 1.66 dB gap that is never explained. The caption says all ablations use the same iterations and setting, but it doesn't say that setting matches the protocol used for the Table 1 retrain. If it does, the baseline is drastically undertrained; if it doesn't, the Table 1 comparison is confounded by compute. Also, row 6 of Table 3 (their full method) shows 33.19 dB on GoPro, which is 1.02 dB below the 34.21 reported in Table 1 for the same method—so the ablation numbers are not on the same scale as the main results. That doesn't invalidate the within-ablation comparison, but it means the absolute numbers can't be cross-referenced. The stress-test's suspicion that row 1's 32.45 may have been copied from MIMO-UNet+ (whose GoPro PSNR is exactly 32.45) is plausible and worth checking.\n\nAlso, the abstract's \"state-of-the-art\" claim is overbroad; on RealBlur-J they are behind HI-Diff and DeepRFT+. The text is more careful, but the abstract should be toned down.\n\nThese are fixable. The authors need to clarify the ablation training protocol, reconcile or relabel the baseline, and ideally release code or at least detailed settings. Given the promising out-of-domain gains, I'd send this to review, but the referee should press on the ablation baseline.","headline":"A new frequency-domain kernel fusion module with promising out-of-domain numbers, but an unexplained 1.66 dB gap between the ablation baseline and the retrained backbone makes the core contribution hard to evaluate.","tokens_in":12616,"tokens_out":3533,"would_cite":false,"duration_ms":28429,"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":"A frequency-domain Transformer that fuses an estimated blur-kernel prior improves blind deblurring and out-of-domain generalization.","keywords":["blind image deblurring","kernel prior","frequency domain","Transformer","out-of-domain generalization","Frequency Integration Module","deconvolution theorem","image restoration"],"falsifier":"Retrain the FFTFormer backbone under the paper's own stage-II and stage-III schedules without any kernel prior, measure GoPro PSNR, and then add FIM with the same schedule; if the FIM-augmented model does not beat that matched baseline by roughly the claimed margin, the kernel-prior effect is not established. A simpler check is to explain the 1.66 dB gap between the 34.11 dB retrained backbone and the 32.45 dB ablation baseline.","tokens_in":11491,"feed_emoji":"🌀","tokens_out":6452,"duration_ms":50787,"temperature":0.7,"pith_summary":"This paper tries to establish that blind image deblurring networks generalize better when they are given an explicit estimate of the blur kernel, and that the most effective way to inject that estimate is to fuse it in the frequency domain rather than by simple channel concatenation. The authors build a kernel estimation module that predicts pixel-wise kernels from the blurry image and a Frequency Integration Module that combines those kernel features with image features through a frequency-domain attention-like operation before each block of a frequency-based Transformer backbone. They report state-of-the-art results on GoPro, HIDE, and RealBlur, with the largest gains appearing on out-of-domain data: 34.21 dB on GoPro, 32.08 on HIDE, 36.59 on RealBlur-R, and 29.01 on RealBlur-J. The importance of the claim is that kernel prior is content-independent, so it may be a practical route to making learned deblurring less dataset-bound.","feed_headline":"Kernel prior boosts blind deblurring where models usually fail","feed_subtitle":"Adding an estimated blur kernel in the frequency domain lifts GoPro-trained deblurring by up to 1.13 dB on RealBlur.","key_machinery":"The Frequency Integration Module (FIM) is the load-bearing mechanism. It projects image and kernel features into query, key, and value triples and computes Frequency Attention as $FA(Q,K,V)=L(F^{-1}(F(Q)\\odot F(K)))\\odot V$, where $F$ is the Fourier transform, $F^{-1}$ its inverse, $L$ layer normalization, and $\\odot$ element-wise multiplication. This mimics the deconvolution theorem, that blur in the spatial domain is pointwise multiplication in the frequency domain, and lets the kernel act as a frequency-domain mask on the image feature. The module is inserted before each encoder and decoder block of a frequency-domain Transformer backbone, at multiple kernel-feature scales, with a residual connection that preserves the original feature; a U-Net-style kernel estimator trained by reblurring the sharp image with the predicted kernel provides the pixel-wise kernel prior.","core_discovery":"On its own terms, the paper's central discovery is that the blur kernel, estimated as a spatially varying map and converted to a frequency-domain conditioning signal, supplies information about the degradation that is independent of image content. When fused by Frequency Attention, the Hadamard product of Fourier-transformed image query and kernel key, followed by inverse transform and element-wise modulation of the value, the kernel prior makes a frequency-domain Transformer both sharper on in-domain data and substantially more robust on out-of-domain benchmarks. The authors show the fusion works best in both encoder and decoder, at multiple scales, with a residual connection inside the module, and that giving the module ground-truth kernels improves results further, indicating the module itself is sound and the bottleneck is kernel estimation accuracy.","pith_inferences":["If the mechanism is as content-independent as claimed, a similar frequency-domain kernel-fusion block should transfer to other restoration tasks that have an underlying degradation kernel, such as super-resolution or atmospheric turbulence correction; this is a direct testable extension the paper does not run.","The pattern of gains, largest on RealBlur-R and smaller on GoPro, suggests the practical value of kernel priors may lie less in pushing the in-domain ceiling and more in stabilizing deployment across camera systems; a controlled cross-dataset sweep with multiple training sets could quantify this.","One internal comparison needs scrutiny before attributing the gains: Table 3's no-kernel baseline is reported at 32.45 dB, while the same backbone retrained by the authors is stated at 34.11 dB; whether the ablation baseline is the identical network under identical training determines how much of the improvement is actually due to the kernel prior."],"forward_implications":["On the GoPro-trained model, adding the kernel prior raises PSNR from 34.11 to 34.21 dB on GoPro, from 31.62 to 32.08 dB on HIDE, and from 35.46 to 36.59 dB on RealBlur-R, with SSIM at least matching on GoPro and improving on HIDE, RealBlur-R, and RealBlur-J.","Out-of-domain generalization improves more than in-domain performance, which is the pattern the authors predict from the kernel prior being content-independent.","Ablations show that both the residual connection and multi-scale encoder-decoder insertion are required; removing the residual connection turns FIM from helpful into harmful.","When true kernels replace estimated ones, PSNR rises further on all four synthetic test sets, so the framework can absorb better kernel estimators without architectural change.","The full model adds moderate cost over the backbone: 24.4 million parameters and 250.2 GFLOPs versus 16.6 million and 131.8 GFLOPs for the retrained FFTFormer."],"supporting_citations":[{"why":"Supplies the kernel estimation approach: a U-Net-style network predicts pixel-wise blur kernels supervised by an L1 reblurring loss against the sharp image.","marker":"[11]"},{"why":"Provides the frequency-domain Transformer backbone, training loss, and the retrained baseline whose numbers the method is compared against.","marker":"[12]"},{"why":"GoPro is the training dataset and the primary in-domain evaluation benchmark.","marker":"[21]"},{"why":"RealBlur-R and RealBlur-J provide the real-world out-of-domain test sets where the largest gains are reported.","marker":"[26]"},{"why":"HIDE is the additional out-of-domain synthetic benchmark used to demonstrate generalization.","marker":"[28]"},{"why":"Sets the training and evaluation protocol, GoPro training and PSNR/SSIM metrics, that the paper aligns with.","marker":"[37]"}],"fun_headline_variants":["Frequency-domain kernel prior sharpens blind deblurring generalization","Blur kernel in frequency domain improves blind deblurring robustness","Kernel prior fusion lifts blind deblurring on unseen domains","Frequency attention with kernel prior beats SOTA blind deblurring","Kernel prior: frequency trick for blind deblurring generalization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the 32.45 dB no-kernel baseline in the ablation study is the same backbone trained under the same conditions as the FIM-augmented model; if that baseline is weaker than the backbone's true retrained performance, the reported gains from the kernel prior are partly an artifact of the comparison.","fun_headline_variants_meta":{"raw":{"variants":["Frequency-domain kernel prior sharpens blind deblurring generalization","Blur kernel in frequency domain improves blind deblurring robustness","Kernel prior fusion lifts blind deblurring on unseen domains","Frequency attention with kernel prior beats SOTA blind deblurring","Kernel prior: frequency trick for blind deblurring generalization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000291,"raw_usage":{"total_tokens":1651,"prompt_tokens":848,"completion_tokens":803,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":464,"completion_tokens_details":{"reasoning_tokens":713}},"tokens_in":464,"tokens_out":803,"duration_ms":5680,"temperature":1.0,"reasoning_tokens":713,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:43:17.579409+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the FFTFormer backbone under the paper's own stage-II and stage-III schedules without any kernel prior, measure GoPro PSNR, and then add FIM with the same schedule; if the FIM-augmented model does not beat that matched baseline by roughly the claimed margin, the kernel-prior effect is not established. A simpler check is to explain the 1.66 dB gap between the 34.11 dB retrained backbone and the 32.45 dB ablation baseline.","supporting_citations":[{"cited_title":"Koalanet: Blind super-resolution using kernel-oriented adaptive local adjustment","cited_arxiv_id":null,"evidence_quote":"Supplies the kernel estimation approach: a U-Net-style network predicts pixel-wise blur kernels supervised by an L1 reblurring loss against the sharp image."},{"cited_title":"Efficient frequency domain-based trans- formers for high-quality image deblurring","cited_arxiv_id":null,"evidence_quote":"Provides the frequency-domain Transformer backbone, training loss, and the retrained baseline whose numbers the method is compared against."},{"cited_title":"Real-world blur dataset for learning and benchmarking de- blurring algorithms","cited_arxiv_id":null,"evidence_quote":"RealBlur-R and RealBlur-J provide the real-world out-of-domain test sets where the largest gains are reported."}],"review_version":1}