{"id":"38c93a07-db39-49b4-b1c9-10afe377300b","arxiv_id":"1908.03826","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DeblurGAN-v2 uses a feature pyramid network and a relativistic GAN to deblur images 10 to 100 times faster than prior methods while staying competitive in quality.","lead":"A new AI model, DeblurGAN-v2, removes motion blur from single photos faster and with quality matching or beating prior methods. The key innovation is a flexible architecture that can be tuned for either high quality or very fast, lightweight operation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's undefined 'linear image subset' and unspecified timing setup make the headline speed/quality comparison unverified; the central claim should stay conditional until the protocol is specified.","rationale":"The paper presents a coherent empirical architecture study with a plausible design and an ablation that supports the incremental contributions. No formal proof is claimed, so correctness risk lives in the experimental comparison. The reader's weakest assumption matches my own: the undefined 'linear image subset' and unstated timing hardware are the most load-bearing ambiguities because Table 1 is the sole evidence for the headline 10-100x speed and close-to-state-of-the-art quality claims. If prior numbers were not recomputed on the same subset and hardware, the comparison could be materially unfair in either direction. The missing error bars further weaken the small metric differences. I do not see an internal inconsistency strong enough to recommend rejection; the promised code release and the detailed ablation are genuine supporting evidence. The appropriate verdict remains CONDITIONAL, pending a clear protocol description and an independent rerun of Table 1.","tokens_in":12556,"tokens_out":4148,"duration_ms":44835,"concrete_test":"Ask the authors to define the 'linear image subset' and release the exact evaluation script; then, using the official released checkpoints, run DeepDeblur, SRN, DeblurGAN, and the three DeblurGAN-v2 variants on the full GoPro test set with identical preprocessing, resolution, batch size 1, and the same GPU (e.g., Tesla P100), recording PSNR/SSIM and wall-clock latency. If MobileNet-DSC's quality drops or its latency advantage over all nearest competitors falls below 10x under this protocol, the central claim should be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—DeblurGAN-v2 with MobileNet-DSC is 10-100x faster than nearest competitors while maintaining close to state-of-the-art deblurring quality—rests almost entirely on Table 1. The caption states 'All models were tested on the linear image subset' but never defines this subset, states whether the prior methods' numbers were recomputed on it, or explains how the subset was selected. If the subset is easier than the full GoPro test set, or if prior reported PSNR/SSIM values were taken from other papers that used different preprocessing, the quality comparison is not valid. Similarly, the timing column reports 'averaged running time per image measured on a single GPU' without naming the GPU model, the image resolution, batch size, or whether prior timing numbers were re-measured under identical conditions. The paper also does not provide error bars or multiple runs, so small differences such as the SSIM gap between DeblurGAN-v2 and SRN (0.934 vs 0.932) may not be meaningful. These ambiguities directly affect the abstract's speed and quality claims, making the comparison protocol the most load-bearing assumption. This is a support gap rather than an internal inconsistency, so the correct response is to demand clarification and an independent rerun before accepting the quantitative claims.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DeblurGAN-v2, a GAN-based single-image motion deblurring architecture whose generator incorporates a Feature Pyramid Network (FPN) for multi-scale feature aggregation and can be paired with different backbones (Inception-ResNet-v2, MobileNet, MobileNet-DSC) to trade off quality against efficiency. The discriminator is a double-scale relativistic LSGAN. The authors report PSNR/SSIM, inference time, and FLOPs on GoPro, Kohler, DVD, and Lai datasets, an ablation study, and an extension to general image restoration. The central claims are that the Inception-ResNet-v2 variant reaches state-of-the-art or close-to-state-of-the-art quality while being faster than prior deep deblurring methods, and that the MobileNet-DSC variant is 10-100x faster than nearest competitors while maintaining competitive quality, enabling near real-time video deblurring.","tokens_in":12774,"tokens_out":4096,"duration_ms":40014,"significance":"If the reported comparisons hold up, the contribution is practically significant: it brings FPN into image restoration, demonstrates a flexible quality-efficiency trade-off through interchangeable backbones, and provides a compact model (about 4 MB for MobileNet-DSC) with large inference speedups. The release of code, models, and data is a concrete strength that increases reproducibility. The ablation study in Table 5 consistently shows improvements from each added component, which supports the internal design choices. However, the quantitative significance is currently bounded by an under-specified comparison protocol in Table 1 and the absence of uncertainty estimates, so the headline claims are not yet fully verifiable from the manuscript alone.","major_comments":[{"comment":"The caption states that all models were tested on the 'linear image subset' of GoPro, but this subset is never defined, and the paper does not state whether the numbers for prior methods were recomputed on this subset or taken from the original publications. This ambiguity is load-bearing for the central speed and quality claims, because a non-representative subset or inconsistent preprocessing would invalidate the comparisons. Please define the subset, its size and selection rule, and state exactly how each baseline's PSNR/SSIM and timing numbers were obtained.","section":"Sec. 4.2, Table 1"},{"comment":"The inference-time comparison reports 'averaged running time per image measured on a single GPU' without naming the GPU model, the image resolution, the batch size, or whether all timings were re-measured under identical conditions. Since the paper's headline contribution is the 10-100x speedup, these details are essential; please specify the hardware and protocol, and ideally report the timing variance across runs.","section":"Sec. 4.2, Table 1"},{"comment":"No error bars, multiple-run statistics, or significance tests are reported, and small metric differences are used to support ranking claims (e.g., SSIM 0.934 vs 0.932 for DeblurGAN-v2 (Inception-ResNet-v2) vs SRN on GoPro, and the 'almost tie' on Kohler). Please add variance information or clearly restrict the claims to differences that exceed noise, especially because the paper states that DeblurGAN-v2 (Inception-ResNet-v2) 'outperforms SRN in SSIM' on GoPro by only 0.002.","section":"Sec. 4.2, Tables 1 and 2"}],"minor_comments":[{"comment":"The text refers to 'Table 6' when presenting the DVD results, but the DVD comparison is actually Table 3; the reference should be corrected.","section":"Sec. 4.4"},{"comment":"The ablation row 'Removing perceptual loss (replace 0.5 with 0 in LG)' is inconsistent with the loss definition LG = 0.5*Lp + 0.006*LX + 0.01*Ladv; the text should say that the perceptual-loss coefficient 0.006 is set to zero. Additionally, the surrounding text says that adding either MSE or perceptual loss benefits training, but only the removal of perceptual loss is shown in the table; consider adding the corresponding MSE-removal row or rephrasing.","section":"Sec. 3.3 and Table 5"},{"comment":"The claim that MobileNet-DSC is 'over 100 times faster than [33, 45]' is not supported by Table 1: compared with SRN, the speedup is 1.6s / 0.04s = 40x. Please adjust the wording to distinguish the 100x speedup relative to DeepDeblur from the 40x speedup relative to SRN.","section":"Sec. 1, bullet list"},{"comment":"There are several typos: 'The author then generated' should be 'The authors then generated', 'join enhancement' should be 'joint enhancement', 'atural images' should be 'natural images', and Table 1 uses 'FLOPS' while the text uses 'FLOPs' elsewhere; these should be corrected.","section":"Sec. 3.4 and Sec. 4.7"},{"comment":"The phrase 'near real-time video frame deblurring, for 25-fps videos' based on 0.04s per image assumes that per-frame processing time is the only bottleneck; please state that other video-level costs are not accounted for, or remove the real-time implication if it is not measured.","section":"Sec. 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical contribution with released code, so the unresolved protocol details are fixable rather than fatal. The main caution for the editor is that Table 1 is currently the linchpin of the paper's most visible claims, and it is not independently verifiable from the text alone; the authors should be required to specify the 'linear image subset' and the timing protocol before the paper is accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"DeblurGAN-v2 is a genuinely useful engineering contribution. The hook is the FPN generator with plug-and-play backbones; the MobileNet-DSC variant, at 4 MB and 0.04 s/image, is a legitimate step toward on-device deblurring. I think the core claims are probably right, but the paper's headline evidence—Table 1—is under-specified enough that the quantitative speed/quality comparison should stay conditional until the protocol is clarified.\n\nWhat is actually new: the first use of Feature Pyramid Networks in deblurring, the double-scale RaGAN-LS discriminator, and a systematic ablation showing each component adds a small consistent gain. The experiments are broad: GoPro, Kohler, DVD, a subjective study on Lai, and a general-restoration extension. The comparison against the authors' own DeblurGAN is fair, and the ablation starts from that baseline, so no circularity. The code and models are promised, which makes verification possible.\n\nThe soft spots are real but not fatal. Table 1's caption says all models were tested on the 'linear image subset' but never defines that subset or states whether prior numbers were recomputed on it. If those PSNR/SSIM values came from other papers, the comparison is apples-to-oranges; if they were re-run, we need the subset criteria. The timing column lacks the GPU model, image resolution, batch size, and whether prior methods were timed under identical conditions. There are no error bars; the SSIM gap between Inception-ResNet-v2 and SRN (0.934 vs 0.932) is small enough that it might not reproduce. The real-time video claim extrapolates from per-image latency without accounting for video pipeline overhead. And the manuscript has at least two table cross-reference errors (Section 4.4 cites Table 6 for DVD results; Section 4.6 cites Table 6 for the ablation), which is minor but indicates the final revision needs a careful pass.\n\nWho is this for? Practitioners who want a fast deblurring model with a tunable speed/quality knob, and researchers working on efficient GAN architectures. It deserves a serious referee: the architecture and the speed-quality trade-off are worth reporting even if the exact numbers need re-validation. My recommendation is to send to peer review and insist that the authors clarify the comparison protocol, specify the hardware and subset definition, and add at least one additional training run for the main comparison.\n\nReading group: maybe—not because the idea is weak, but because the evaluation-protocol ambiguities make a good discussion case for how benchmarks get reported.","headline":"DeblurGAN-v2 is a solid engineering contribution—FPN generator plus flexible backbones and a genuinely fast lightweight variant—but Table 1's undefined 'linear image subset' and unspecified timing protocol keep the headline quantitative claims conditional.","tokens_in":13371,"tokens_out":5698,"would_cite":true,"duration_ms":52226,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","68T45"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that one GAN architecture, built around an FPN generator with swappable backbones and a double-scale relativistic discriminator, matches or beats the best prior deblurring models in quality while running 10-100 times…","keywords":["single-image deblurring","generative adversarial network","feature pyramid network","relativistic GAN","least-squares GAN","real-time video deblurring","mobile vision","image restoration"],"falsifier":"Re-run SRN, DeepDeblur, and DeblurGAN on the exact same linear GoPro subset using the same GPU and timing protocol and compare wall-clock times and PSNR or SSIM; if DeblurGAN-v2 with MobileNet-DSC is less than 10 times faster or its quality gap widens beyond the reported margin, the central efficiency claim fails.","tokens_in":12317,"feed_emoji":"⚡","tokens_out":6397,"duration_ms":62508,"temperature":0.7,"pith_summary":"DeblurGAN-v2 is a generative adversarial network for removing motion blur from a single image. The paper claims it produces images as sharp as the best prior deblurring models, matching or beating them on PSNR, SSIM, and human preference, while being far cheaper to run. With a heavy backbone it competes with the strongest published methods; with a light backbone it runs 10-100 times faster than its nearest rivals, at 0.04 seconds per 1280x720 image and a 4 MB model, which the authors say makes real-time video deblurring feasible. The source of the gain is a Feature Pyramid Network in the generator, which aggregates multi-scale features in one pass instead of feeding multiple image scales through separate streams, plus a double-scale relativistic discriminator and a hybrid loss.","feed_headline":"Fast deblurring GAN hits 10-100x speedup with near-top quality","feed_subtitle":"A 4-MB variant processes a 1280x720 frame in 0.04 s, opening the door to real-time video deblurring.","key_machinery":"The central object is the Feature Pyramid Network (FPN) inserted as the generator's core: a bottom-up convolutional backbone produces features at five scales, a top-down pathway up-samples the semantically rich low-resolution features, lateral connections add high-resolution detail, and the resulting multi-scale maps are up-sampled and concatenated into one tensor. This replaces the multi-stream, multi-scale input pyramids used by prior deblurring CNNs with a single forward pass, which is what yields the efficiency gain. Around it sits a double-scale discriminator with one global branch on the full image and one local patch branch, using a relativistic least-squares (RaGAN-LS) loss that estimates how much more realistic a real image is than a fake one, and a generator loss $L_G = 0.5 L_p + 0.006 L_X + 0.01 L_{adv}$ combining MSE, VGG19 perceptual, and adversarial terms. The backbone is swappable; the paper demonstrates Inception-ResNet-v2, MobileNet, and a depthwise-separable-convolution MobileNet variant.","core_discovery":"On the paper's own terms, the discovery is that an FPN-based generator turns deblurring into a plug-and-play trade-off: the same adversarial training recipe, with a relativistic least-squares discriminator that looks at both whole images and local patches, yields top-tier quality when the backbone is powerful and near-real-time speed when the backbone is a compact MobileNet. The authors report that the Inception-ResNet-v2 variant ranks at the top of GoPro SSIM and beats competitors in subjective tests on real blurry photos, while the MobileNet-DSC variant keeps SSIM close to DeblurGAN and DeepDeblur while running over one hundred times faster than the slower baselines and eleven times faster than DeblurGAN. They state that MobileNet-DSC is, to their knowledge, the only deblurring method that simultaneously delivers reasonably high performance and that level of inference efficiency, and they show the same generator improves a joint restoration benchmark as well.","pith_inferences":["The same FPN generator could be dropped into other image-to-image restoration problems such as denoising, deraining, or super-resolution, where multi-scale aggregation is currently done with pyramid inputs; the speed advantage suggests similar orders-of-magnitude gains there, though the paper only sketches this extension.","If the 0.04 seconds per frame figure holds end-to-end, deblurring moves from an offline batch operation to a mobile or embedded camera pipeline, which would make blur-robust preview, video stabilization, and real-time correction viable in consumer devices.","The unspecified 'linear image subset' of GoPro leaves a concrete open task: a standardized re-benchmark of all methods on the full test set with a common timing protocol would determine whether the reported speed-quality frontier is robust or partly an artifact of the evaluation subset.","Because perceptual quality improves even where PSNR does not, the relativistic GAN training suggests that human-perceived sharpness and standard fidelity metrics can diverge for real-world motion blur, so a perceptual metric tailored to blur may better capture what users actually see."],"forward_implications":["With the Inception-ResNet-v2 backbone, DeblurGAN-v2 matches or beats the strongest published deblurring methods on quality metrics, achieving top SSIM on GoPro, a near-tie with SRN on Kohler, and the best subjective score on Lai, while using about one-fifth of SRN's inference time.","With the MobileNet backbone, DeblurGAN-v2 outperforms the video-deblurring baselines WFA and DVD in the single-frame setting on the DVD dataset while running at least 17 times faster.","With MobileNet-DSC, the model reaches 0.04 seconds per image and a 4 MB size, which the authors state enables near-real-time deblurring of 25-fps video frames; they claim it is the only method so far combining reasonably high quality with that efficiency.","The ablation shows each component, FPN, global discriminator, RaGAN-LS, and MSE loss, adds measurable PSNR and SSIM gains, with FPN contributing the most, so the design choices reinforce one another.","Because the architecture is backbone-agnostic, the same framework extends to general image restoration: on a synthetic dataset with blur, noise, compression, and upscaling artifacts, the Inception-ResNet-v2 variant outperforms DeblurGAN on PSNR and SSIM."],"supporting_citations":[{"why":"Baseline DeblurGAN; the paper's direct predecessor and the efficiency comparison point, since DeblurGAN-v2 with MobileNet-DSC is 11 times faster, and the starting point for the ablation study.","marker":"[21]"},{"why":"Supplies the GoPro dataset and the DeepDeblur multi-scale baseline; its PSNR, SSIM, and runtime are central comparison targets.","marker":"[33]"},{"why":"Scale-Recurrent Network baseline, the closest quality competitor in PSNR and SSIM and the main object of the efficiency comparison, including the compact-SRN control experiment.","marker":"[45]"},{"why":"Introduces the Feature Pyramid Network module that the generator adopts to aggregate multi-scale features in a single forward pass.","marker":"[27]"},{"why":"Provides the relativistic discriminator concept used to build the RaGAN-LS loss for stable and fast training.","marker":"[16]"},{"why":"Provides the least-squares GAN objective that the paper wraps into the relativistic loss.","marker":"[30]"},{"why":"Supplies the MobileNetV2 backbone used for the light-weight and depthwise-separable-convolution variants.","marker":"[40]"},{"why":"Supplies the Inception-ResNet-v2 backbone used for the high-performance variant.","marker":"[44]"},{"why":"Defines the perceptual loss on VGG19 features used as the content term in the generator loss.","marker":"[15]"}],"fun_headline_variants":["DeblurGAN-v2: 100x faster deblurring, near-top quality","Real-time video deblurring: DeblurGAN-v2 hits 10-100x speedup","FPN-based GAN deblurs 10-100x faster with mobile backbones","Speed-accuracy knob: DeblurGAN-v2 from real-time to SOTA","DeblurGAN-v2: orders-of-magnitude faster, still competitive"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison is fair: every method's speed and quality numbers come from the same 'linear image subset' of GoPro under the same timing protocol, but the paper does not specify what that subset is or whether the prior published numbers were recomputed on it.","fun_headline_variants_meta":{"raw":{"variants":["DeblurGAN-v2: 100x faster deblurring, near-top quality","Real-time video deblurring: DeblurGAN-v2 hits 10-100x speedup","FPN-based GAN deblurs 10-100x faster with mobile backbones","Speed-accuracy knob: DeblurGAN-v2 from real-time to SOTA","DeblurGAN-v2: orders-of-magnitude faster, still competitive"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000347,"raw_usage":{"total_tokens":1942,"prompt_tokens":1028,"completion_tokens":914,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":644,"completion_tokens_details":{"reasoning_tokens":796}},"tokens_in":644,"tokens_out":914,"duration_ms":9062,"temperature":1.0,"reasoning_tokens":796,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:59:57.343112+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run SRN, DeepDeblur, and DeblurGAN on the exact same linear GoPro subset using the same GPU and timing protocol and compare wall-clock times and PSNR or SSIM; if DeblurGAN-v2 with MobileNet-DSC is less than 10 times faster or its quality gap widens beyond the reported margin, the central efficiency claim fails.","supporting_citations":[{"cited_title":"Deblurgan: Blind motion deblurring using conditional adversarial networks","cited_arxiv_id":null,"evidence_quote":"Baseline DeblurGAN; the paper's direct predecessor and the efficiency comparison point, since DeblurGAN-v2 with MobileNet-DSC is 11 times faster, and the starting point for the ablation study."},{"cited_title":"Deep Multi-scale Convolutional Neural Network for Dy- namic Scene Deblurring","cited_arxiv_id":null,"evidence_quote":"Supplies the GoPro dataset and the DeepDeblur multi-scale baseline; its PSNR, SSIM, and runtime are central comparison targets."},{"cited_title":"Scale-recurrent network for deep image deblurring","cited_arxiv_id":null,"evidence_quote":"Scale-Recurrent Network baseline, the closest quality competitor in PSNR and SSIM and the main object of the efficiency comparison, including the compact-SRN control experiment."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the least-squares GAN objective that the paper wraps into the relativistic loss."}],"review_version":1}