{"id":"0f7d04d2-1a6b-48c5-ad94-ef961841fa65","arxiv_id":"2511.10806","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"A ViT+FFT-ReLU cascade for image deblurring is reported as state-of-the-art, but its PSNR/SSIM gains over the ViT alone are negligible (0.00–0.03 dB) and the method's two-stage interface is never specified.","lead":"A team at the University of Dhaka combines a popular transformer network (Restormer) with a frequency-domain FFT-ReLU deblurring module and claims state-of-the-art results. The paper's own table shows the new pipeline matches Restormer exactly on two of four benchmarks and beats it by only 0.02–0.03 dB on the others, while a small, author-selected human survey is the main evidence for superiority.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed SOTA gain is unsupported: 'Ours' numerically equals the Restormer preprocessor on HIDE/GoPro and differs by ≤0.03 dB elsewhere, and Algorithm 1's input is never tied to the ViT stage.","rationale":"The reader's weakest assumption—Algorithm 1's input is never tied to the ViT output—is a real and central flaw, and I agree it is load-bearing. My emphasis is slightly different: the paper's own Table 1 quantitatively undermines the headline claim even before the pipeline ambiguity is resolved, because 'Ours' is numerically identical to Restormer on two of four benchmarks and differs by at most 0.03 dB on the other two. Without error bars, paired tests, or a clear description of what the FFT stage receives, there is no reproducible evidence that the frequency module adds any value. The pipeline ambiguity is what prevents the reader from even checking whether the 0.02–0.03 dB differences are real, so both concerns point to the same conclusion: the central claim is unsubstantiated. I do not see a need to move the verdict; the REJECT verdict stands. I did not identify any independent support that would offset this: no machine-checked proofs or reproducible artifacts were available, and the human preference study (76 raters, 10 author-selected image sets, no statistical test) cannot rescue the quantitative claim. The critique is directed at the argument, not the authors: the missing specification could in principle be fixed by code release, and the concrete test would settle it.","tokens_in":12196,"tokens_out":3607,"duration_ms":34671,"concrete_test":"Obtain the released code and run the pipeline in two instrumented configurations on the RealBlur-R test set: (A) Algorithm 1 receives the original blurred image as B, with the ViT output used only to initialize/restrict the kernel; (B) Algorithm 1 receives the Restormer deblurred output as B. Report PSNR/SSIM with per-image paired differences and bootstrap confidence intervals, comparing against Table 1's 36.21/0.957 and Restormer alone (36.19/0.957). If (B) reproduces 36.21, the FFT stage is processing a non-blurred image and violates the forward model; if neither configuration beats Restormer by more than ~0.05 dB with non-overlapping CIs, the claimed superiority is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that the ViT+FFT-ReLU cascade is superior to state-of-the-art—requires two things: the FFT stage must receive a valid blurred image satisfying the single-convolution forward model B = S * k (Algorithm 1, §3.2.1), and the cascade must measurably improve on the ViT stage. Neither is established. In Table 1, 'Ours' is identical to Restormer on HIDE (31.22/0.942) and GoPro (32.92/0.961), and exceeds it by only 0.02–0.03 dB PSNR on RealBlur-R/J, with no error bars or significance tests. A 0.0125 dB average PSNR gain is not evidence of superiority. The methodological ambiguity in §3 and Fig. 1 explains why: Algorithm 1's input B is used as the blurred observation for kernel estimation, but Fig. 1 labels that input 'Blurred Image With Reduced Kernel.' If B is the Restormer output, it is not a blurred image consistent with B = S * k, so the estimated kernel is meaningless. If B is the original blurred image, the ViT output is not used in the FFT stage, and the claimed tight coupling is absent. The paper never states which case holds, so the reported numbers cannot be attributed to the proposed integration. The 'first hybrid' claim is also weakened by prior CNN-Transformer and frequency-selection hybrids cited in [4,19,33], but the decisive flaw is the missing causal link between the two stages.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage image deblurring pipeline: a Vision Transformer (specifically Restormer) first preprocesses the blurred image and is claimed to 'narrow down the blur kernel values', and then an FFT-based blind/non-blind deconvolution stage with ReLU sparsity (the authors' own WACV 2025 method) restores the sharp image. The authors claim this is the first tightly coupled hybrid of spatial attention and frequency-domain sparsity, and report superior PSNR/SSIM over state-of-the-art on GoPro, HIDE, RealBlur, and Kohler, supplemented by a human preference study and runtime/memory analysis.","tokens_in":12638,"tokens_out":2770,"duration_ms":26090,"significance":"If the claimed integration were real and the gains were reproducible, a lightweight frequency-domain postprocessor that improves a strong ViT baseline would be a useful practical contribution, especially given the reported small added runtime and memory cost. The paper also provides code, a human evaluation, and explicit discussion of the limitations of pixel-wise metrics. However, the central quantitative claim is not supported by the reported numbers, and the connection between the two stages is not specified, so the contribution is currently unsubstantiated. The novelty claim of being the first spatial-frequency hybrid is also weakened by prior work cited in the paper itself.","major_comments":[{"comment":"The claim that the proposed method 'generates superior results in terms of PSNR and SSIM' is not supported by Table 1. The 'Ours' row is identical to Restormer on HIDE (31.22/0.942) and GoPro (32.92/0.961), and exceeds it by only 0.03 dB PSNR on RealBlur-R and 0.02 dB on RealBlur-J. The average PSNR gain over Restormer is 0.0125 dB. No error bars, standard deviations, or significance tests are provided, and the gains are far below typical benchmark noise. This does not establish superiority.","section":"Table 1, Section 4.3.1"},{"comment":"The pipeline's input to Algorithm 1 is never specified. Algorithm 1's input B is used as the blurred observation in a single-convolution forward model B = S * k. Figure 1 labels the intermediate signal 'Blurred Image With Reduced Kernel', implying B is the ViT output. If so, the ViT output is not a blurred image satisfying the forward model, making the kernel estimate meaningless. If instead B is the original blurred image, the ViT output is not used in the FFT stage, and the claimed 'tightly coupled' integration disappears. The manuscript must state which input is used and provide an ablation; without this, the reported numbers cannot be attributed to the proposed integration.","section":"Section 3, Figure 1, Algorithm 1"},{"comment":"The mechanism by which the ViT 'narrows down the blur kernel values' is not described or validated. The ViT is a standard deblurring network trained to output sharp images; there is no explanation of how its output constrains the PSF estimate in Algorithm 1. This is a load-bearing step for the claimed synergy, and no experiment isolates its effect.","section":"Section 3.1"},{"comment":"The paper itself concedes that quantitative gains are 'small' and repositions the contribution as 'practically superior' based on the human preference study. However, the human study in Section 4.3.3 reports only mean ranks from a small set of 76 participants over 10 image sets, with no error bars, no statistical test, and no information about participant selection or image diversity. A 9.67/10 mean for 'Ours' with no variance measure does not substantiate perceptual superiority over methods scoring 3.84–5.91.","section":"Section 4.3.4"},{"comment":"The claim that this is 'the first to present a tightly coupled hybrid model that combines both [spatial and frequency]' is contradicted by prior work cited in the paper, including [4] (CNN–Transformer combination), [19] (frequency selection for deblurring), and [33] (inverted residual Fourier transformation for deblurring). The novelty statement should be narrowed to the specific ViT-plus-FFT-ReLU architecture and should acknowledge existing spatial–frequency deblurring hybrids.","section":"Section 1 and Related Works (Section 2)"}],"minor_comments":[{"comment":"Formatting error: '36.190.957' should read '36.19 0.957'. The caption says 'highlighted' but no rows are bolded. Also, the 'Average' column uses inconsistent arithmetic precision.","section":"Table 1"},{"comment":"The algorithm uses undefined variables 'weight' and 'opts.xk_iter'. Procedure 1's findM and threshold_pxpy_v1 are mentioned but not defined, making the algorithm not self-contained.","section":"Algorithm 1"},{"comment":"The figure caption is confusing: subfigures (a) and (b) are labeled 'Vision Transformer' and 'FFT with ReLU Sparsity', but the text in the figure seems mismatched with the caption. Please clarify the flow and labels.","section":"Figure 1"},{"comment":"The experimental setup only states that 'Restormer is trained on GoPro' but does not give training details, hyperparameters, or how the FFT stage parameters (λ_ftr, λ_grad, λ_tv, λ_L0, weight_ring, etc.) are set. This makes reproduction difficult.","section":"Section 4.2"},{"comment":"Runtime analysis reports total time as the sum of the two stages, but no comparison is made to a state-of-the-art end-to-end model other than Restormer. The claim that FFT-ReLU scales with 'near-linear memory' needs a plot or reference.","section":"Section 4.3.5"}],"recommendation":"reject","confidential_remarks":"The FFT-ReLU module is essentially the authors' own WACV 2025 paper [1] with three overlapping authors. The present manuscript reuses it as a module and evaluates the combination on the same benchmarks where the ViT baseline already achieves the reported numbers. The quantitative gain over Restormer is negligible, and the input ambiguity in Section 3/Fig. 1 means the causal contribution of the proposed integration is not established. I see no clear path to acceptance without substantial new experiments and re-framing, which goes beyond a minor revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a serial pipeline of two existing pieces — Restormer as a front-end and your own FFT-ReLU blind deconvolution (WACV 2025) as a back-end. The paper's own Table 1 shows 'Ours' exactly matching Restormer on HIDE (31.22/0.942) and GoPro (32.92/0.961), and exceeding it by only 0.02–0.03 dB on the RealBlur sets. That is not evidence of a working integration; it is baseline noise. There are no error bars or significance tests.\n\nWhat is good: the write-up is clear in places, the algorithms are given in pseudocode, and the runtime analysis is honest — the hybrid roughly doubles inference time and still uses 23 GB because of the transformer. The FFT-ReLU stage itself is lightweight, and the authors at least attempted a human preference study.\n\nThe real problem is that the method as described cannot be evaluated. Algorithm 1 assumes its input is a blurred image obeying B = S * k, but Figure 1 labels the input to that stage 'Blurred Image With Reduced Kernel' — which sounds like the Restormer output. Those are incompatible. If you feed a deblurred image into blind deconvolution, the forward model is violated and the kernel estimate is meaningless. If you feed the original blur, the ViT contributes nothing to the FFT stage. The paper never says which case is used, so the reported numbers cannot be attributed to the proposed integration.\n\nThe 'first tightly coupled hybrid' claim also doesn't survive contact with the paper's own references ([4], [19], [33] are CNN-Transformer or frequency-selection hybrids). And the human study — 76 raters, ten author-selected image sets, no statistical test, average score 9.67/10 — is not a meaningful perceptual evaluation.\n\nBottom line: the central claims are unsupported by the presented evidence. This should not go to full peer review in its current form; it is a desk-reject candidate. It could serve as a teaching example of how a superficial cascade can produce a misleading paper, but I would not cite it.","headline":"A cascade of Restormer and the authors' own FFT-ReLU that shows no meaningful gain over the baseline and never specifies how the two stages actually connect.","tokens_in":13184,"tokens_out":2659,"would_cite":false,"duration_ms":23692,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A dual-domain deblurring pipeline pairs a Vision Transformer with FFT-ReLU sparsity and claims state-of-the-art results on RealBlur-R and HIDE.","keywords":["image deblurring","Vision Transformer","FFT-ReLU","frequency-domain sparsity","blind deconvolution","dual-domain restoration","PSNR/SSIM evaluation","human visual preference"],"falsifier":"Take a synthetic blurred image generated by convolving a sharp image with a known kernel, feed the Transformer's output into Algorithm 1, and compare the estimated kernel to the ground-truth kernel; if the estimate does not resemble the true kernel, the FFT stage is not doing the deblurring the paper attributes to it.","tokens_in":12083,"feed_emoji":"🖼️","tokens_out":4142,"duration_ms":39541,"temperature":0.7,"pith_summary":"This paper claims that image deblurring improves when a Vision Transformer and a frequency-domain FFT-ReLU deconvolution stage are joined in a single pipeline, rather than used separately. The Transformer is meant to narrow the possible blur kernel, and the FFT-ReLU stage then enforces frequency sparsity to suppress ringing and restore fine edges. On the RealBlur-R and HIDE benchmarks the authors report the best PSNR/SSIM among the compared models, and a human-preference survey strongly favors their output. If true, this would establish a new dual-domain paradigm that bridges spatial attention and frequency-domain priors.","feed_headline":"Vision Transformer plus FFT sparsity tops deblurring benchmarks","feed_subtitle":"Two-stage design narrows the blur kernel in space, then cleans up edges in frequency; human viewers preferred it over prior models.","key_machinery":"The load-bearing machinery is the FFT-ReLU sparsity prior: an iterative blind-deconvolution loop that alternates latent-image estimation via L0 gradient minimization with fast total variation regularization and PSF estimation in the Fourier domain, followed by a non-blind ringing-artifact removal stage. The Vision Transformer acts as a preprocessor meant to shrink the space of possible blur kernels before this frequency-domain optimization runs.","core_discovery":"In the author's own terms, the contribution is the first tightly coupled hybrid model that unifies spatial-domain attention with frequency-domain sparsity for deblurring: a Vision Transformer preprocesses the blurred image to reduce kernel ambiguity, then FFT-based blind and non-blind deconvolution with ReLU sparsity restores the sharp image. The authors claim this two-phase design outperforms existing deblurring models on RealBlur-R and HIDE by PSNR and SSIM, matches the best results on GoPro and RealBlur-J, and produces images that human viewers prefer by a wide margin.","pith_inferences":["A testable extension: replace the Transformer preprocessor with a cheaper kernel-reduction method, such as a small CNN or a hand-crafted edge detector, and compare results; if performance holds, the advantage may come from the frequency prior rather than long-range attention.","The same dual-domain recipe could transfer to other restoration tasks such as denoising or super-resolution, where high-frequency detail recovery is also limited by purely spatial models.","If the Transformer truly only narrows the kernel space, then the method should handle spatially varying blur gracefully; analyzing the estimated PSF across image regions would clarify whether the underlying uniform-kernel model is realistic."],"forward_implications":["If the pipeline works as claimed, spatially-aware Transformers can be paired with cheap frequency-domain deconvolution to sharpen edges without greatly increasing memory use.","The near-linear memory scaling of the FFT-ReLU stage suggests the hybrid can be extended to high-resolution images where pure Transformer deblurring is impractical.","The human-preference result implies pixel-based metrics understate perceptual gains; future deblurring evaluations could combine metrics with structured human assessment.","Reported performance on HIDE, which features non-uniform human motion blur, indicates the frequency prior is not limited to uniform blur, though the paper does not provide an ablation isolating this factor."],"fun_headline_variants":["Hybrid ViT and FFT cleans up blur better than rivals","Two-stage deblurring: ViT then FFT-ReLU wins on metrics","ViT+FFT deblurring beats state of the art on benchmarks","Spatial attention meets frequency sparsity for sharper images","Dual-domain model excels at restoring sharp photos"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The claimed spatial–frequency synergy assumes that the Vision Transformer's output can still be treated as a blurred image with a reduced blur kernel when it enters the blind-deconvolution stage, even though that output is itself a deblurred image.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid ViT and FFT cleans up blur better than rivals","Two-stage deblurring: ViT then FFT-ReLU wins on metrics","ViT+FFT deblurring beats state of the art on benchmarks","Spatial attention meets frequency sparsity for sharper images","Dual-domain model excels at restoring sharp photos"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00063,"raw_usage":{"total_tokens":2717,"prompt_tokens":683,"completion_tokens":2034,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":427,"completion_tokens_details":{"reasoning_tokens":1942}},"tokens_in":427,"tokens_out":2034,"duration_ms":12052,"temperature":1.0,"reasoning_tokens":1942,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T22:20:34.930123+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a synthetic blurred image generated by convolving a sharp image with a known kernel, feed the Transformer's output into Algorithm 1, and compare the estimated kernel to the ground-truth kernel; if the estimate does not resemble the true kernel, the FFT stage is not doing the deblurring the paper attributes to it.","supporting_citations":[],"review_version":1}