{"id":"75b87950-a19f-439b-8c8a-a962b24b9f8b","arxiv_id":"2412.01559","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"AHFNet dynamically weights four handpicked high-pass kernels (Sobel and temporal gradients) to extract sharpening features, reaching 33.25 dB PSNR on GOPRO with roughly one-sixth the training memory of heavier models.","lead":"AHFNet restores blurry videos by predicting a lightweight combination of a few high-pass filters that add back edges and fine detail, claiming top accuracy among low-memory models with dramatically faster inference. The design trades a small accuracy loss against much lower training memory, which could make video deblurring practical on phones and embedded devices.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 1 is false as stated under the standard -3 dB cutoff definition, and the implementation never enforces the non-negative coefficient assumption; the paper's claimed high-pass guarantee for the predicted kernel lacks support.","rationale":"The reader's verdict is CONDITIONAL, which I think is right. The empirical comparison in Tables 1 and 2 is the main evidence, and the Proposition 1 issue does not by itself invalidate the measured PSNR/SSIM. However, the paper's narrative is built around the claim that the adaptive kernel is guaranteed high-pass; Proposition 1 is the only stated justification, and it is unproved in the main text and likely false under the standard definition. Even more directly, the implementation described in Eq. 5 does not satisfy the proposition's non-negativity assumption. Thus the theoretical foundation is not secure. I agree with the reader's identification of Proposition 1 as a weak spot, but I would frame the issue as 'the theorem is false as stated and not applicable to the implemented model,' rather than only 'cutoff definition ambiguous.' The empirical results still justify a conditional acceptance pending the proposition fix and stronger baselines/error bars. No change to the reader's verdict is needed.","tokens_in":14085,"tokens_out":5180,"duration_ms":47866,"concrete_test":"Two checks would settle the concern. (1) Analytical: derive the magnitude response of the sum of two ideal high-pass filters with cutoffs fc1=0.1 and fc2=0.9 and compute the -3 dB cutoff; if it exceeds fc1, Proposition 1 is false as stated. (2) Empirical: using the released code, extract the generated kernel k_t and the predicted coefficients alpha_t for a set of GOPRO test frames. Test whether all alpha_t are non-negative; compute the DC gain (sum of kernel values) and the -3 dB cutoff of k_t via DFT. If alpha_t contains negative values or the DC gain is non-zero, the claimed high-pass guarantee does not hold for the deployed model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.3 introduces Proposition 1, which asserts that any non-negative linear combination of high-pass filters is high-pass with cutoff not greater than the smallest basis cutoff fc1, and then uses it to claim that the dynamically predicted kernel k_t in Eq. 5 is guaranteed high-pass. Two load-bearing problems arise. First, the cutoff statement is not true under the usual -3 dB definition: for two ideal high-pass filters with cutoffs 0.1 and 0.9, the sum has response 0 below 0.1, 1 between 0.1 and 0.9, and 2 above 0.9; normalized to the passband maximum, the -3 dB point occurs at 0.9, not at or below 0.1. The proof is deferred to supplementary material and is not present in this version, so the claim cannot be checked. Second, Proposition 1 requires non-negative alpha_i, but Eq. 5 defines alpha via an unconstrained network G and no non-negativity activation is reported anywhere in the architecture description or training details. If alpha can be negative, the linear combination of high-pass bases need not have zero DC response, so even the weaker 'zero DC gain' notion of high-pass is not guaranteed. These issues do not necessarily overturn the empirical PSNR/SSIM results, but they remove the theoretical foundation for the paper's core claim that AHFNet explicitly constrains the network to high-frequency extraction. Without a corrected proposition or a direct measurement of the learned kernels' frequency responses, the interpretability and mechanism claims are unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AHFNet, a video deblurring network that explicitly extracts high-frequency features by dynamically predicting mixing coefficients over a fixed set of high-pass basis kernels (spatial Sobel filters and temporal difference kernels). The generated kernels are convolved with the input video, along with their 90-degree rotations, and the resulting features are fed into a recurrent deblurring module. The authors report state-of-the-art performance among methods with similar training memory footprint on GOPRO and DVD, with 33.25 dB PSNR on GOPRO, 461 GMACs, and 0.144 s inference time, and they attribute the gains to the explicit high-pass structure, supported by Proposition 1.","tokens_in":14393,"tokens_out":5776,"duration_ms":42775,"significance":"If the results hold, AHFNet offers a practical efficiency/performance trade-off for video deblurring, with a simple and interpretable design that is substantially cheaper in training memory than transformer/deformable-convolution baselines while achieving competitive PSNR/SSIM. Strengths include the clear architectural description, the frequency-subband analysis in Figure 6, ablations over the number of extraction paths (Table 4) and basis kernel choices (Table 5), and the promise of released code. The main caveats are the theoretical guarantee (Proposition 1) and the statistical robustness of the small performance differences.","major_comments":[{"comment":"The proposition's claim that the cutoff frequency of a non-negative linear combination of high-pass filters is not greater than the smallest basis cutoff fc1 is false under the standard -3 dB definition. For two ideal high-pass filters with cutoffs 0.1 and 0.9, the combined response is 0 below 0.1, 1 between 0.1 and 0.9, and 2 above 0.9; normalized to the passband maximum of 2, the -3 dB crossing occurs at 0.9, not at or below 0.1. The proof is deferred to supplementary material and is not available in this version. Furthermore, Proposition 1 requires non-negative coefficients αi, but Eq. (5) defines α via an unconstrained network G, and no non-negativity activation (e.g., ReLU or softmax) is described in the architecture or training details. Without non-negativity, the generated kernel kt need not have zero DC response, so even the weak 'zero DC gain' notion of high-pass is not guaranteed. This undermines the central interpretability claim that AHFNet explicitly constrains the network to high-frequency extraction. Please correct the proposition or its proof, enforce or justify non-negativity of α, and/or directly measure the frequency responses of learned kernels to empirically support the high-pass claim.","section":"Section 3.3, Proposition 1 and Eq. (4)-(5)"},{"comment":"The headline improvement of 0.33 dB PSNR from 'integrating spatial and temporal gradients' is computed against the '+ RGB×2' control, which lacks any high-pass basis and simply repeats the early module with RGB features. The appropriate ablation for isolating the effect of the high-pass constraint is the comparison against the equally complex '3D Conv' and 'Naive Kernels' variants in Table 3, where the gains are 0.07 dB and 0.09 dB, respectively. As reported, the abstract and Section 4.4 overstate the contribution of the explicit high-frequency structure, conflating the basis-kernel choice with the overall module design. Please report the comparison against the learned-kernel variant as the primary ablation or revise the claim accordingly.","section":"Section 4.4, Table 5 and Abstract"},{"comment":"No variance or number of runs is reported for PSNR/SSIM, and several key differences are small (e.g., Table 3: AHFNet 32.89 vs 3DConv 32.82; Table 4: N=4 vs N=6, 0.18 dB). Without error bars or significance tests, the claimed improvements over the closest baselines are not statistically supported. Please report mean and standard deviation over at least three runs, or otherwise demonstrate that the differences exceed run-to-run variability.","section":"Tables 1-5"}],"minor_comments":[{"comment":"There is a stray period and line break after 'including both forward and backward propagation.' in the memory footprint definition; please remove the extra punctuation.","section":"Section 4.1"},{"comment":"The notation for the temporal basis kernels is ambiguous: the first kernel is written as [[0],[−1],[1]] and the second as [[1],[−1/2],[−1/2]]; please clarify the frame ordering (e.g., t−1, t, t+1) and the Gram-Schmidt derivation, which appears to start from [[1],[−1],[0]] rather than from the first basis kernel.","section":"Section 3.3"},{"comment":"The symbols ∇x, ∇2x, and ∇tx are used without formal definition; please define them (e.g., first-order spatial gradient, Laplacian, temporal gradient) in the text or in a table.","section":"Section 4.4"},{"comment":"The y-axis label 'Relative MSE Values' and the subband labels are difficult to read; consider enlarging the font and using a clearer axis description.","section":"Figure 6"},{"comment":"The claim of 'optimal trade-off' is based on a single point per method; consider adding error bars or noting that memory values are taken from official implementations.","section":"Section 4.2, Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The paper's empirical contribution is plausible and the resource analysis is useful, but the theoretical proposition in Section 3.3 is incorrect as stated and the non-negativity of the predicted coefficients is not enforced. This is a serious issue for the interpretability narrative, though it may not invalidate the empirical gains. The authors should be asked to either fix the proposition, provide a corrected definition/proof, or reframe the contribution as an empirical design choice, and to add direct frequency-response measurements of the learned kernels. Additionally, the 0.33 dB claim against the RGB×2 baseline should be repositioned. I would not recommend rejection if these issues are addressed in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know up front: the empirical story here is credible and useful. AHFNet hits 33.25 dB PSNR on GOPRO with 461 GMACs and 0.144 s per 1280x720 frame, beating every model with a comparable training memory footprint and running roughly 35x faster than ERDN. The architecture is simple enough to reimplement, the ablation in Table 5 shows a 0.33 dB gain from adding the Sobel + temporal gradient extraction over an 'RGB x 2' control, and the code is on GitHub. If you need a low-memory deblurring baseline, this is a serious candidate.\n\nWhat is actually new is the specific combination: predicting coefficients over fixed high-pass bases, rotating those bases, and aggregating via absolute-value summation. None of the building blocks are new, but the package is, and the efficiency numbers are good.\n\nNow the soft spot, and it is load-bearing for the interpretability claim. Proposition 1 says any non-negative linear combination of high-pass filters is high-pass with cutoff no greater than the smallest basis cutoff. That is not true under the standard -3 dB definition. The stress-test example (two ideal high-pass filters with cutoffs 0.1 and 0.9, sum normalized) puts the cutoff at 0.9, not 0.1. The proof is deferred, and Eq. 5 doesn't enforce non-negative alphas, so even the zero-DC version of 'high-pass' isn't guaranteed. This doesn't overturn the PSNR/SSIM results, but it removes the theoretical foundation for saying the network is explicitly constrained to high-frequency extraction. The authors should either fix the proposition, drop it, or measure the learned kernels' frequency responses directly.\n\nOther issues are minor. No error bars on the PSNR/SSIM, the 'RGB x 2' baseline is weak, and the claimed low-memory comparison omits heavier models like ShiftNet and RVRT from the tables. None of these are unusual for the field.\n\nBottom line: this paper is for practitioners who want a fast, cheap video deblurrer and for people interested in frequency-based inductive biases. It deserves a serious refereeing: the empirical work is solid, the code is out, and the theory can be repaired or honestly downgraded. I'd take it to peer review and ask for a corrected proof, a non-negativity constraint, and error bars.\n\n—","headline":"Solid low-budget video deblurring with a real 35x speedup, but the paper's theoretical guarantee for the high-pass kernel is unsupported and likely wrong as stated.","tokens_in":14916,"tokens_out":3292,"would_cite":true,"duration_ms":27477,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A network that predicts adaptive high-pass kernels from Sobel and temporal gradient bases achieves 33.25 dB PSNR on GOPRO, the best among models with similar training memory, and runs 35x faster than ERDN.","keywords":["video deblurring","high-frequency extraction","adaptive kernel prediction","high-pass filters","efficient inference","spectral bias","kernel prediction networks","GOPRO-DVD benchmarks"],"falsifier":"Evaluate the frequency response of a non-negative linear combination of two ideal high-pass filters with cutoffs at 0.1 and 0.9; under the standard half-power cutoff definition, the combined response reaches its half-power point above 0.9, which would directly falsify Proposition 1's 'no greater than the smallest cutoff' claim. A separate experiment could train AHFNet with a kernel that is provably high-pass by construction and compare against the current coefficient-prediction module; equal performance would show the guarantee is not what drives the accuracy.","tokens_in":13843,"feed_emoji":"🎥","tokens_out":11407,"duration_ms":84480,"temperature":0.7,"pith_summary":"Blur suppresses the high-frequency content of video, yet standard deep network components—self-attention and ordinary convolutions—tend to weaken or under-learn exactly those frequencies. The paper's thesis is that explicit high-frequency extraction is the missing ingredient, so it builds AHFNet around a module that predicts four mixing coefficients over a fixed set of high-pass basis kernels: two Sobel filters and two temporal differencing kernels. The generated kernel is convolved with the blurry input window, and the extracted feature maps feed a recurrent deblurring backbone, generalizing classical unsharp masking. On GOPRO the full model reaches 33.25 dB PSNR and 0.9439 SSIM, the best among models with comparable training memory, while requiring about 461 GMACs and 0.144 seconds per 1280×720 frame—roughly 35 times faster than ERDN. The authors report that the simplest version of the module, spatial Sobel plus temporal gradients, adds 0.33 dB PSNR over an equal-complexity control, which they take as evidence that the explicit high-pass structure, not extra capacity, drives the improvement.","feed_headline":"High-pass kernel prediction makes video deblurring fast and accurate","feed_subtitle":"Four mixing coefficients over Sobel and temporal gradients hit 33.25 dB with minimal training memory.","key_machinery":"The adaptive high-frequency extraction module $\\mathcal{H}$ is the load-bearing component. It holds $M=4$ fixed 3D high-pass basis kernels $\\{\\tilde{k}_j\\}$—two normalized 3×3 Sobel filters (horizontal and vertical) and two temporal differencing kernels, one of which is Gram–Schmidt orthogonalized—and a small coefficient generator $G$ that reads the blurry input window and outputs mixing coefficients $\\{\\alpha_{t,j}\\}$. The generated kernel $k_t=\\sum_j \\alpha_{t,j}\\tilde{k}_j$ is convolved with the input, along with its 90°-rotated counterpart, and the outputs $h_t$, $h'_t$, and $|h_t|+|h'_t|$ form the extracted high-frequency feature set. Proposition 1 is the mechanism that connects this construction to the high-pass guarantee: any non-negative linear combination of high-pass filters is claimed to remain high-pass, which is why the network can predict only $M$ coefficients instead of the full kernel. The module is placed at the head of the network, with each of $N$ paths feeding a shared recurrent extraction module that produces the residual added to the input frame.","core_discovery":"The central claim is that deblurring can be reformulated as explicit high-frequency recovery: a blurry frame $x_t$ plus a learned residual $\\mathcal{L}(x_t, \\mathcal{H}(\\{x_i\\}_{i=t-l}^{t+l}))$, where $\\mathcal{H}$ is a module whose kernels are constrained to be high-pass. The constraint is enforced by construction: the module predicts non-negative mixing coefficients $\\{\\alpha_{t,j}\\}_{j=1}^M$ and forms $k_t = \\sum_j \\alpha_{t,j} \\tilde{k}_j$ from four fixed basis kernels—two normalized 3×3 Sobel kernels and two temporal difference kernels—then applies $k_t$, its 90°-rotated version, and the summed absolute values to the input window. Proposition 1 is invoked to guarantee that $k_t$ remains high-pass. The paper reports that this module lifts PSNR by 0.33 dB over a matched 'RGB×2' control, outperforms variants with randomly generated high-pass bases by 0.8 dB, and yields the best PSNR/SSIM among models trained within a comparable memory budget: 33.25 dB / 0.9439 on GOPRO and 33.19 dB / 0.9400 on DVD.","pith_inferences":["An implication the paper leaves implicit is that the same module could transfer to single-image deblurring, super-resolution, or denoising, where spectral bias also suppresses fine details; nothing in the module is video-specific beyond the temporal basis kernels.","If Proposition 1's cutoff claim is not valid under the standard -3 dB definition, the empirical results would still stand, but the explanation would shift from an enforced spectral guarantee to the inductive bias of fixed gradient bases—a distinction future ablations could test by comparing against a provably high-pass kernel of the same capacity.","A testable extension would be spatially varying coefficients: the current module predicts one mixing coefficient per kernel per frame, so the adaptive kernel is global; allowing per-pixel or per-patch coefficients could adapt the filter to local blur anisotropy at a modest computational cost.","The success of coefficient prediction over four bases hints that larger basis sets, such as oriented edge filters at multiple scales, could extend the same efficiency-accuracy trade-off, since only the coefficient dimension grows."],"forward_implications":["If the central claim is right, the deblurring head does not need expensive transformer or deformable-convolution components: a handful of constrained convolutional paths suffices, cutting training memory to about one-sixth of models like FGST and inference to 0.144 s per frame.","The coefficient-prediction design reduces the learnable weights of the dynamic kernel from the full kernel size to $M=4$, so the module adds little in parameters or GMACs (461.35 vs 410.73 with no extraction paths) while improving PSNR by 0.33 dB.","Because the basis kernels are fixed and interpretable, the resulting model behaves as a learned generalization of unsharp masking, suggesting that frequency-based inductive biases can substitute for black-box architectural complexity in restoration tasks.","The module's gains are concentrated in mid-to-high frequency subbands, where the paper shows lower relative MSE than the 3D-Conv and Naive-Kernel variants, so the benefit is specifically about high-frequency recovery.","Scaling the number of extraction paths $N$ from 0 to 6 monotonically improves PSNR on GOPRO (32.56 to 32.89 dB), indicating the module is a composable building block."],"supporting_citations":[{"why":"Supplies the GOPRO benchmark and the blur accumulation model in Eq. 1, which frames deblurring as low-pass filtering and motivates recovering lost high frequencies.","marker":"[25]"},{"why":"Provides the DVD dataset and the DBN baseline used for comparison in the low-budget model evaluation.","marker":"[38]"},{"why":"Establishes the spectral bias of neural networks, the problem the explicit high-pass module is designed to counter.","marker":"[33]"},{"why":"Documents that self-attention behaves as a low-pass filter, motivating explicit high-pass extraction.","marker":"[31]"},{"why":"Dynamic filter networks, the direct-kernel-prediction alternative compared against in Section 4.3.","marker":"[14]"},{"why":"Basis prediction networks, the coefficient-over-basis scheme that the module adapts to high-pass bases.","marker":"[46]"},{"why":"ERDN, the deformable-convolution baseline over which the paper claims a 35x speedup with superior PSNR and SSIM.","marker":"[15]"},{"why":"MMP-RNN, the strongest low-budget baseline whose PSNR and runtime are compared in Table 1.","marker":"[44]"},{"why":"Source of the random high-pass basis kernels used as a baseline in the kernel-choice ablation of Section 4.4.","marker":"[42]"},{"why":"Classical unsharp masking, the conceptual precursor that Eq. 2 generalizes to learned, inter-frame high-frequency extraction.","marker":"[7]"}],"fun_headline_variants":["Adaptive high-pass kernels sharpen video deblurring efficiently","Explicit high-frequency recovery speeds up video deblurring","Predict high-pass kernels for fast video deblurring","Efficient video deblurring with adaptive high-pass kernels","High-pass kernel prediction boosts deblurring accuracy and speed"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is Proposition 1, stated in Section 3.3 without proof in the main text: any non-negative linear combination of high-pass filters is itself a high-pass filter with cutoff no greater than the smallest basis cutoff; if that frequency guarantee is wrong, the generated kernels are not high-pass by construction and the reported improvements would rest on the network learning useful filters on its own.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive high-pass kernels sharpen video deblurring efficiently","Explicit high-frequency recovery speeds up video deblurring","Predict high-pass kernels for fast video deblurring","Efficient video deblurring with adaptive high-pass kernels","High-pass kernel prediction boosts deblurring accuracy and speed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000464,"raw_usage":{"total_tokens":2321,"prompt_tokens":953,"completion_tokens":1368,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":1283}},"tokens_in":569,"tokens_out":1368,"duration_ms":8791,"temperature":1.0,"reasoning_tokens":1283,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:17:48.217597+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the frequency response of a non-negative linear combination of two ideal high-pass filters with cutoffs at 0.1 and 0.9; under the standard half-power cutoff definition, the combined response reaches its half-power point above 0.9, which would directly falsify Proposition 1's 'no greater than the smallest cutoff' claim. A separate experiment could train AHFNet with a kernel that is provably high-pass by construction and compare against the current coefficient-prediction module; equal performance would show the guarantee is not what drives the accuracy.","supporting_citations":[{"cited_title":"Deep video deblurring for hand-held cameras","cited_arxiv_id":null,"evidence_quote":"Provides the DVD dataset and the DBN baseline used for comparison in the low-budget model evaluation."},{"cited_title":"On the spectral bias of neural networks","cited_arxiv_id":null,"evidence_quote":"Establishes the spectral bias of neural networks, the problem the explicit high-pass module is designed to counter."},{"cited_title":"Dynamic filter networks","cited_arxiv_id":null,"evidence_quote":"Dynamic filter networks, the direct-kernel-prediction alternative compared against in Section 4.3."},{"cited_title":"Basis prediction networks for effective burst denoising with large kernels","cited_arxiv_id":null,"evidence_quote":"Basis prediction networks, the coefficient-over-basis scheme that the module adapts to high-pass bases."},{"cited_title":"Erdn: Equivalent receptive field deformable network for video deblurring","cited_arxiv_id":null,"evidence_quote":"ERDN, the deformable-convolution baseline over which the paper claims a 35x speedup with superior PSNR and SSIM."},{"cited_title":"Efficient video deblurring guided by motion magnitude","cited_arxiv_id":null,"evidence_quote":"MMP-RNN, the strongest low-budget baseline whose PSNR and runtime are compared in Table 1."},{"cited_title":"A generalized unsharp masking algorithm","cited_arxiv_id":null,"evidence_quote":"Classical unsharp masking, the conceptual precursor that Eq. 2 generalizes to learned, inter-frame high-frequency extraction."}],"review_version":1}