{"id":"02855620-f28d-4dd2-98c7-637e78e3c3d1","arxiv_id":"2501.18403","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A reduced Restormer variant with doubled attention heads and extra augmentations reaches parity with the original model on RealBlur and UHDM deblurring benchmarks at 18.4% fewer parameters.","lead":"This paper tweaks the Restormer image-deblurring model, cutting its parameter count by about 18 percent and adding color and blur augmentations plus a frequency-based training loss. Tests on three deblurring benchmarks show the smaller model roughly matches the original while training about 5 hours faster, though the small quality differences lack error bars.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The efficiency claim is confounded: the improved model changes architecture, training schedule, augmentations, and loss simultaneously, so the 18.4% parameter reduction is never isolated as the cause of maintained performance.","rationale":"The reader's weakest assumption is exactly that the baseline and improved model differ only by intended modifications, and Section VI.B/VII show they do not: training schedules, augmentations, and loss functions all change. My stress-test identifies the same confound as the single most load-bearing concern. This does not require changing the reader's CONDITIONAL verdict: the paper's central efficiency claim is plausible but not yet isolated, and a single controlled ablation could settle it. No additional independent concern outweighs this one; the marginal numerical gaps and missing error bars strengthen the concern but do not move the verdict to rejection because the paper does provide code and datasets for verification. I agree with the reader's assessment, so the verdict remains unchanged.","tokens_in":12995,"tokens_out":2268,"duration_ms":22314,"concrete_test":"Run a controlled ablation: train the reduced-architecture model (Figure 3) using exactly the baseline training pipeline from Section VI.B (the 1-GPU progressive schedule with batch sizes 8/4/4/2/1/1, horizontal and vertical flips only, L1 loss only, 300K iterations) and evaluate on UHDM and RealBlur-R/J. If this reduced-architecture model matches the reproduced baseline within noise, the architecture is performance-neutral; if it is significantly worse, the maintained performance in Table VI depends on the training changes rather than the architecture. A complementary check is to train the original baseline architecture with the improved pipeline; if it matches the improved model, the gains are training-only.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that reducing layer count by 30% and parameters by 18.4% maintains or improves deblurring quality. To support that claim, the comparison must isolate architecture as the causal factor. It does not. The reproduced baseline (Section VI.B) is trained with progressive schedule {(128,8), (160,4), (192,4), (256,2), (320,1), (320,1)}, L1 loss only, and no color/Gaussian/perspective augmentations. The improved model (Section VII) is trained with a different schedule {(128,8), (160,6), (192,4), (256,2), (320,2), (384,1)}, additional color jitter, Gaussian blur, brightness/contrast, perspective transforms, and the composite loss Ltotal = Lpixel + 0.1 Lfreq. The ablations in Section VII compare the full changed system against the reproduced baseline; no single-component ablation isolates the architectural modification. Moreover, the reported gains are marginal: UHDM PSNR improves from 21.260 to 21.359 dB (0.099 dB) and RealBlur-R improves from 33.685 to 33.997 dB, but the provided checkpoint already scores 33.998 dB on RealBlur-R. Without error bars or multi-seed runs, these differences are within plausible noise. Consequently, the observed performance could be entirely due to the improved training pipeline, with the architecture reduction being performance-neutral or even harmful while the extra augmentations and frequency loss compensate. The stated conclusion that 'the reduction in parameters and total layers made the model more efficient while maintaining its representational power' is therefore not established by the presented evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents an empirical study of Restormer for high-resolution image motion deblurring. The authors reproduce a baseline Restormer, fine-tune the authors' checkpoint on RealBlur-R/J, and propose an 'improved' variant that reduces the number of transformer blocks/layers and total parameters, doubles the attention heads per stage, adds color jitter/Gaussian blur/brightness-contrast/perspective augmentations, and trains with a composite loss consisting of pixel L1 plus a Fourier-magnitude loss. The paper reports PSNR, SSIM, MAE, LPIPS, and DeltaE on RealBlur-R, RealBlur-J, and UHDM, plus training and inference time comparisons. The main claim is that the improved model maintains or improves deblurring performance while reducing model complexity by 18.4% and total layers by 30%.","tokens_in":13308,"tokens_out":6627,"duration_ms":58736,"significance":"If the central claim were established, the paper would provide a practically useful data point: a smaller Restormer variant with comparable high-resolution deblurring quality and lower computational cost. The authors are transparent about their training schedules and about the difficulties of reproducing the original Restormer results, which is a strength. However, the evidence does not currently isolate the architectural modification as the cause of the observed performance; the improved model differs from the baseline simultaneously in architecture, training schedule, augmentations, and loss function, and the reported performance differences are small relative to plausible training variability. The paper also provides no code or trained model in the submission beyond an unverified repository URL, so the reproducibility of the claimed 18.4% parameter reduction depends entirely on a precise architectural description that is not fully given in the text.","major_comments":[{"comment":"The central efficiency claim is not supported because the comparison is confounded. The reproduced baseline (Section VI.B) is trained with the progressive schedule {(128,8),(160,4),(192,4),(256,2),(320,1),(320,1)}, L1 loss only, and no color or geometric augmentations beyond flips, whereas the improved model (Section VII) is trained with {(128,8),(160,6),(192,4),(256,2),(320,2),(384,1)}, the composite loss L_total = L_pixel + 0.1 L_freq, and additional color jitter, Gaussian blur, brightness/contrast, and perspective transforms. The 'ablation studies' in Section VII compare the full changed system against the reproduced baseline; no experiment varies the architecture alone. Therefore the maintenance or improvement in Tables III, V, and VI cannot be attributed to the 18.4% parameter reduction and 30% layer reduction claimed in the abstract.","section":"Section VII; Section VI.B"},{"comment":"The quantitative evidence for improvement is statistically weak. All reported values come from a single training run, with no error bars, multiple seeds, or significance tests. On UHDM the improved model's PSNR is 21.359 dB versus 21.323 dB for the provided checkpoint and 21.260 dB for the reproduced model; on RealBlur-R the improved model (33.997 dB) is essentially identical to the checkpoint (33.998 dB). Differences of roughly 0.1 dB on a single seed are within the range of random initialization and training variability, especially given the paper's own acknowledgment that data-split details are unknown. The authors should report means and standard deviations over multiple seeds, or a paired significance test, before claiming that the improved model maintains or improves performance.","section":"Section VI.D; Tables III, V, VI"},{"comment":"The comparison against the published Restormer results is not valid as presented. The rows labeled 'Results in the paper' were obtained with 8 GPUs, batch sizes up to 64, and the schedule {(160,40),(192,32),(256,16),(320,8),(384,8)}, while the reproduced model is trained on 1 GPU with batch sizes up to 8 and a different schedule. Section VI.B itself states that lack of information about the train/validation split could explain the deviations. Consequently, neither the reproduced nor the improved model can be said to 'maintain' the original Restormer's performance; the comparison conflates architecture with training protocol and data handling.","section":"Section VI.B; Tables II and IV"},{"comment":"The architectural modification is not specified with enough precision to verify the claimed parameter reduction. The text states that the number of transformer blocks and layers was decreased and that attention heads per stage were doubled, but it does not give the exact per-level block counts, head counts, channel counts, or the refinement-stage configuration of the improved model. Without these numeric details, the 18.4% parameter reduction and 30% layer reduction cannot be independently reproduced from the manuscript, and the efficiency claim is not checkable.","section":"Section IV; Figures 2 and 3"}],"minor_comments":[{"comment":"The text refers to a hard-example threshold of 'PSNR between 20dB and 30dB', but Table I says 'PSNR between 20dB and 3dB'; the definition of Hard Positives/Hard Negatives is also not formalized, and the counts are not obviously comparable across datasets of different sizes.","section":"Table I"},{"comment":"The enumeration of the four experimental configurations is inconsistent: the word 'Second' is used twice and the third/fourth configurations are not clearly introduced, making the experimental setup harder to follow.","section":"Section VI.A"},{"comment":"The phrase 'detailed ablation studies' overstates the content: the section compares the full improved system with the reproduced baseline and shows training curves, but it does not ablate the frequency-loss weight, the augmentation components, or the architectural changes individually.","section":"Section VII"},{"comment":"On UHDM the improved model has a worse DeltaE (4.023) than the reproduced baseline (3.867), which the paper acknowledges in passing. The abstract's blanket statement that the model 'maintains or improves performance' should be qualified to reflect that color fidelity is not improved on this benchmark.","section":"Table VI; Discussion"},{"comment":"The GitHub link is mentioned but the manuscript does not state whether code, model weights, or evaluation scripts are included, nor does it provide a versioned release or license, which weakens the reproducibility statement.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads like a well-documented project report rather than a mature journal paper. The incremental contribution is modest, and the experimental design does not isolate the proposed architectural change, which is the load-bearing claim. The central claim could in principle be repaired by adding single-component ablations (architecture only, training schedule only, augmentations only, loss only) and multi-seed evaluation with error bars. Without those, I would not be comfortable accepting the paper, and even with them the contribution may be better suited to a workshop or a short empirical-study venue unless the authors strengthen the comparison against existing deblurring methods."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper reports a real 18.4% parameter reduction in Restormer and a first UHDM evaluation, but the main efficiency claim is not established because architecture, data augmentation, loss, and training schedule all change at once. The authors are honest about reproduction issues, which I appreciate, but the evidence is not tight enough to support the title claim.\n\nThe useful parts: they reproduce Restormer on RealBlur and show the provided checkpoint deviates from the original paper (33.998 vs 36.190 dB), a legit reproducibility data point. They also benchmark on UHDM, apparently the first Restormer result there, and report inference times and model sizes. The hard-example analysis, while loosely defined, is a reasonable attempt to look beyond aggregate PSNR. Code is linked, though without a commit hash.\n\nThe soft spot is load-bearing. The improved model differs from the reproduced baseline in at least four ways: fewer layers/blocks, doubled attention heads, extra augmentations (color jitter, Gaussian blur, perspective), and an added Fourier magnitude loss. The ablations compare the full changed system against the baseline, not individual components, so the 18.4% parameter reduction is never isolated as the cause of maintained performance. The training schedules also differ: baseline is batch 4/2/1 at 160/256/320; improved is batch 6/2/2/1 at 160/256/320/384. Since the baseline schedule uses smaller batches at higher resolutions and the improved schedule adds a final 384 patch, the comparison is not controlled. The UHDM gain is 21.359 vs 21.323 dB (0.036 dB), almost certainly within run-to-run noise; there are no error bars or multiple seeds. Table I is internally confusing: counts do not sum to totals and the threshold is written \"20dB and 3dB\". The frequency loss is borrowed from Benjdira et al.; calling it \"new\" in the abstract overstates it.\n\nWho this is for: someone working on efficient deblurring might find the UHDM numbers and model-size/inference-time figures useful as a data point, but I would not cite it as evidence that a smaller Restormer works as well. The paper is a competent technical report that needs a controlled experiment before the main claim can be evaluated. I would suggest the authors rerun with a matched training schedule and per-component ablations; then it could be a reasonable workshop paper. As is, I would not send it to a strong venue. Desk reject with an invitation to resubmit after controlled ablations.","headline":"Honest reproduction study, but the efficiency claim is confounded by simultaneous changes to architecture, training schedule, augmentations, and loss.","tokens_in":13892,"tokens_out":2570,"would_cite":false,"duration_ms":23447,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","68U10"],"pacs":[],"model":"deepseek-v4-flash","headline":"A 18.4% smaller Restormer transformer keeps deblurring quality while cutting training time by five hours.","keywords":["image deblurring","Restormer","transformer","motion blur","efficient attention","frequency loss","data augmentation","high-resolution images"],"falsifier":"Run a controlled comparison in which the original Restormer and the simplified model are trained from scratch with identical data splits, augmentations, loss, and the same progressive patch/batch schedule; if the simplified model's PSNR and SSIM then fall below the original's on RealBlur-R or UHDM, the claimed 'maintaining or improving performance' is an artifact of the training recipe rather than the architectural change.","tokens_in":12760,"feed_emoji":"🖼️","tokens_out":4743,"duration_ms":40338,"temperature":0.7,"pith_summary":"This paper argues that the Restormer transformer, a general-purpose image restoration network, is over-parameterized for motion deblurring. By pruning roughly 30% of its layers and transformer blocks while doubling the number of attention heads per stage, the authors obtain a model with 18.4% fewer parameters that trains about five hours faster and matches or slightly exceeds the original's PSNR and SSIM on real-world blur benchmarks. They attribute part of the gain to an enriched training pipeline—color jitter, Gaussian blur, perspective transforms, and a Fourier magnitude loss—that smooths convergence. The result matters because it suggests that careful architectural simplification, not just bigger models, can make transformer-based deblurring practical on limited hardware.","feed_headline":"18.4% smaller Restormer keeps deblurring quality","feed_subtitle":"Doubled attention heads and extra augmentations offset the layer cuts, and training time drops from 28 to 23 hours.","key_machinery":"The core mechanism is the interplay between two Restormer components—Multi-Dconv Head Transposed Attention (MDTA), which computes channel-wise attention with linear complexity, and the Gated-Dconv Feed-Forward Network (GDFN). The modification reduces the number of these blocks across the network but doubles the attention heads per MDTA stage, so the model attends to more subspaces in fewer layers. A Fourier magnitude loss, $L_{freq} = \\frac{1}{N}\\sum_i\\||F(\\hat{I}_i)| - |F(I_i)|\\|_1$ with weighting $\\lambda=0.1$, is added to the pixel L1 loss to preserve high-frequency detail, and the augmented training pipeline (color jitter, Gaussian blur, brightness/contrast, perspective transforms) is intended to bridge the color and geometric variability of real blurred images.","core_discovery":"The central claim is that a deliberately simplified Restormer variant can maintain or improve motion-deblurring quality while being substantially cheaper. The authors reduce the number of transformer blocks and layers in the encoder, latent, decoder, and refinement stages, doubling the attention heads per MDTA stage to preserve representational capacity, and combine this with extra data augmentations and a frequency-domain loss. On RealBlur-R, the improved model reaches 33.997 dB PSNR versus 33.685 dB for their reproduced baseline; on UHDM it reaches 21.359 dB versus 21.260 dB; and it cuts both training time (23 versus 28 hours) and inference time. The improved model also lowers the count of hard negative examples on UHDM from 604 to 598. The authors present this as evidence that many existing architectures may be over-parameterized for their target tasks.","pith_inferences":["The uncontrolled differences in progressive-training schedules between the reproduced baseline and the improved model mean the claimed architectural benefit is not cleanly isolated; a matched-schedule comparison could still confirm it, but the current evidence conflates architecture with training recipe.","If the simplification principle transfers to other restoration tasks such as denoising and super-resolution, it would imply that Restormer's original block counts are not a universal optimum and task-specific pruning could be a general recipe.","The Fourier magnitude loss is a lightweight addition that could be ported to other restoration architectures, but its contribution is not separated from the augmentations in the ablation, so a factorial study would reveal which component drives the smoother convergence.","The UHDM results, where all models score below 21.4 dB PSNR, suggest that high-resolution synthetic blur remains a hard target; a worthwhile extension would test the simplified model on real 4K–6K motion blur rather than only kernel-convolved sharp images."],"forward_implications":["The 18.4% parameter and 30% layer reductions shrink the model to 81.5 MB versus 99.9 MB, making transformer deblurring more feasible on memory-limited devices.","Faster convergence and smoother loss curves suggest the simplified architecture is easier to train, lowering the compute budget needed to reach a given quality.","The improved model's higher PSNR and fewer hard negatives on UHDM indicate the simplification generalizes to very high-resolution blur rather than only the training distribution.","Fine-tuned models on RealBlur degrade sharply on UHDM, revealing a specialization–overfitting tradeoff that the improved model avoids.","The combined architectural and training changes yield a net win even though color augmentation alone did not improve color fidelity on UHDM, pointing to the architecture as the main efficiency driver."],"supporting_citations":[{"why":"Supplies the Restormer architecture, MDTA and GDFN components, and the original progressive-training hyperparameters that the paper modifies.","marker":"[20]"},{"why":"Provides the RealBlur-R and RealBlur-J real-world blur benchmarks used for fine-tuning, reproduction, and final evaluation.","marker":"[10]"},{"why":"Provides the UHDM dataset, the high-resolution motion-blur benchmark on which the improved model posts its best relative gains.","marker":"[23]"},{"why":"Motivates the doubling of attention heads per stage, cited as inspiration for scaling attention capacity under reduced layer count.","marker":"[16]"},{"why":"Supplies the guided frequency loss formulation that the paper adapts into its Fourier magnitude loss term.","marker":"[1]"},{"why":"Provides the GoPro dataset used to retrain the baseline model from scratch for reproducibility checks.","marker":"[9]"}],"fun_headline_variants":["Slimmer Restormer matches deblurring at 82% cost","Restormer cut 18% size, still tops benchmarks","Leaner transformer deblurs faster, equal quality","Efficient transformer trims 18% size, boosts PSNR","Deblur with smaller Restormer: 23h training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that the reproduced baseline and the improved model differ only by the intended modifications, but the baseline was trained with a smaller progressive batch schedule and the RealBlur train/validation split is not known, so the reported performance differences could stem from these uncontrolled factors rather than the architecture.","fun_headline_variants_meta":{"raw":{"variants":["Slimmer Restormer matches deblurring at 82% cost","Restormer cut 18% size, still tops benchmarks","Leaner transformer deblurs faster, equal quality","Efficient transformer trims 18% size, boosts PSNR","Deblur with smaller Restormer: 23h training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000123,"raw_usage":{"total_tokens":1078,"prompt_tokens":899,"completion_tokens":179,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":515,"completion_tokens_details":{"reasoning_tokens":88}},"tokens_in":515,"tokens_out":179,"duration_ms":2189,"temperature":1.0,"reasoning_tokens":88,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T23:36:00.835842+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a controlled comparison in which the original Restormer and the simplified model are trained from scratch with identical data splits, augmentations, loss, and the same progressive patch/batch schedule; if the simplified model's PSNR and SSIM then fall below the original's on RealBlur-R or UHDM, the claimed 'maintaining or improving performance' is an artifact of the training recipe rather than the architectural change.","supporting_citations":[{"cited_title":"Real- World Blur Dataset for Learning and Benchmarking Deblurring Al- gorithms","cited_arxiv_id":null,"evidence_quote":"Provides the RealBlur-R and RealBlur-J real-world blur benchmarks used for fine-tuning, reproduction, and final evaluation."},{"cited_title":"MC- Blur: A Comprehensive Benchmark for Image Deblurring, September","cited_arxiv_id":null,"evidence_quote":"Provides the UHDM dataset, the high-resolution motion-blur benchmark on which the improved model posts its best relative gains."},{"cited_title":"Guided Frequency Loss for Image Restoration","cited_arxiv_id":"2309.15563","evidence_quote":"Supplies the guided frequency loss formulation that the paper adapts into its Fourier magnitude loss term."},{"cited_title":"Deep multi-scale convolutional neural network for dynamic scene deblurring","cited_arxiv_id":null,"evidence_quote":"Provides the GoPro dataset used to retrain the baseline model from scratch for reproducibility checks."}],"review_version":1}