{"id":"eb4438aa-8b28-4e1a-9512-6c89447a382d","arxiv_id":"2608.08066","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"EvBS uses event streams to transfer motion patterns between image patches, synthesizing more diverse blur-sharp pairs and improving domain-adaptive deblurring.","lead":"This paper introduces EvBS, which uses event camera data to generate new blurred versions of sharp patches, giving deblurring models extra training examples in the target domain. Because motion is separated from image content, motion patterns can be moved between scene regions, and the authors report consistent quality gains on real-world deblurring benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains may be inflated because fine-tuning and evaluation use the same target frames, so the claim of improving unseen target domains is not yet established.","rationale":"The paper's central claim is that EvBS enables domain-adaptive deblurring that improves performance on unseen target domains. The most load-bearing requirement is that the reported improvements reflect genuine adaptation rather than leakage from the evaluation data. The paper describes extracting content and motion sources from the target-domain test data, synthesizing blur from those sources, fine-tuning on those pairs, and then reporting metrics on the same datasets. Without an explicit adaptation/evaluation split, the model may overfit to the specific frames used for synthesis, inflating the results. This concern is distinct from the reader's weakest assumption (pseudo-sharp label noise and misalignment), though related: even if pseudo-sharp patches are clean, the evaluation protocol alone can make the numbers optimistic. The paper also tunes key hyperparameters on REVD and then reports REVD in the main table, compounding the leakage. A concrete test that holds out a subset of frames for evaluation would settle whether the improvements generalize. If the gains largely vanish, the central claim fails; if they persist, the method is credible. The reader's verdict of CONDITIONAL is appropriate, and the required condition should include a proper adaptation/evaluation split and hyperparameter selection protocol.","tokens_in":17019,"tokens_out":9663,"duration_ms":109309,"concrete_test":"Hold out a subset of scenes or frames from each dataset (e.g., 30% of EVRB and REVD frames) and use only the remaining frames for source extraction, blur synthesis, and fine-tuning. Then evaluate the adapted model on the held-out frames and compare with the pre-trained baseline. If the PSNR/SSIM gains on the held-out frames are substantially smaller (e.g., below half of the reported gains), the central claim is compromised. Additionally, report results where hyperparameters are tuned on a validation split of one dataset (or on EVRB) and then evaluated on REVD without re-tuning, to rule out hyperparameter leakage. A meaningful control is to fine-tune the same models on pairs synthesized from the held-in frames only and show that the improvements transfer to the held-out frames.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that EvBS improves pre-trained deblurring models on unseen target domains. However, the experimental protocol does not separate the frames used for adaptation from the frames used for evaluation. The Dual Source Extractor draws content sources and motion sources from the target-domain test data, the Blur Conditioning Module synthesizes training pairs from those exact sources, and the pre-trained model is fine-tuned on those pairs and then evaluated on the same dataset (Tables 1 and 2). No scene-level or frame-level split is described. Because the fine-tuning loss directly supervises reconstruction of the pseudo-sharp content sources from blur synthesized from those same patches, the model may learn to invert the specific synthesis for those frames rather than acquire a generalizable target-domain mapping. This risk is amplified by the fact that key hyperparameters (patch ratio r, extrinsic pair count n, and orientation thresholds in Fig. 9) are tuned on REVD, which also appears in the main results, so the REVD numbers are optimistically biased. If the reported 2+ dB gains are partly due to transductive overfitting to the evaluation frames, the paper's core assertion that EvBS 'effectively enhances the robustness of existing deblurring models on unseen testing datasets' is not supported. The reader's concern about label noise and sensor misalignment is legitimate, but the evaluation-protocol issue is more load-bearing because even with perfectly clean pseudo-sharp sources, the current setup cannot distinguish genuine adaptation from memorization.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EvBS, an event-guided blur synthesis framework for domain-adaptive motion deblurring. Given test-domain frames and event streams, a Dual Source Extractor identifies pseudo-sharp content patches and motion-consistent blur patches. A Blur Conditioning Module constructs intrinsic-blur conditions (using the content patch's own event-derived motion) and extrinsic-blur conditions (warping a separate motion source's flow onto the content patch via Algorithm 1), which guide a modified diffusion-based blurring model (ID-Blau) to synthesize blurry/sharp training pairs. Pre-trained event-based and frame-based deblurring models are then fine-tuned on these pairs with the content patches as supervision. The authors report consistent PSNR/SSIM/LPIPS gains across EVRB, REVD, HighREV-3/7/11 for five backbones, and compare favorably to prior domain-adaptive deblurring methods.","tokens_in":17396,"tokens_out":2919,"duration_ms":31713,"significance":"The core idea of using events to explicitly decouple motion from visual content for blur synthesis is novel and addresses a real limitation of RGB-based synthesis methods, which can only re-apply a motion pattern to its original patch. If the evaluation protocol is sound, the method would be a useful contribution to domain-adaptive deblurring, with the additional advantage of natively supporting event-based deblurring models. The paper is well-structured, includes ablations of the main components (Table 3, Fig. 9), a computational cost analysis (Table S1), and an honest limitations section (Supplementary Sec E). However, the evaluation protocol as described does not separate the frames used for adaptation from those used for evaluation, which directly threatens the central claim of improving performance on 'unseen' target domains; hyperparameters are also tuned on REVD, one of the benchmark datasets.","major_comments":[{"comment":"The adaptation procedure uses the target-domain test data itself to build the training set: the Dual Source Extractor draws content and motion sources from the test frames, the blurring model synthesizes pairs from those exact patches, and the pre-trained model is fine-tuned on those pairs and then evaluated on the same datasets (Tables 1 and 2). No scene-level or frame-level split between adaptation and evaluation data is described. The reported gains could therefore reflect transductive overfitting to the evaluation frames rather than generalization to unseen target-domain data. Please adapt on a disjoint subset (e.g., a training split of each target dataset) and evaluate on a held-out split, and report both transductive and disjoint results to quantify this effect.","section":"Section 4.2 / Fig. 2"},{"comment":"Key hyperparameters (patch ratio r, extrinsic pair count n, orientation thresholds) are ablated on REVD, and REVD is also included in the main results in Tables 1 and 2. Tuning hyperparameters on a benchmark and then reporting that same benchmark as a result introduces optimistic bias and does not establish that the chosen values generalize to other datasets. Please fix hyperparameters using a separate validation set (e.g., a held-out split of REVD or a different dataset) and then evaluate on all benchmarks, or report the sensitivity of the main results to these choices across datasets.","section":"Section 4.3 / Fig. 9"},{"comment":"All quantitative results are reported as single runs without standard deviations or significance tests. Given that fine-tuning is stochastic and that source selection and extrinsic-pair selection involve sampling-like procedures, the consistency of the claimed gains cannot be assessed. Please report mean and standard deviation over at least three random seeds, and where feasible a paired significance test for the main comparisons (baseline vs. w/ Ours, and Ours vs. DADeblur).","section":"Tables 1 and 2"},{"comment":"The paper acknowledges that residual blur in selected content patches introduces label noise and that event–image misalignment degrades performance, but these risks are not quantified. Because every synthesized pair inherits the quality of the pseudo-sharp content source, please add an analysis showing how the final deblurring performance varies with the level of residual blur in content sources (e.g., by varying r beyond the single ablation in Fig. 9a) and include a sensitivity study under simulated event–image misalignment.","section":"Supplementary Sec E / Section 4.3"}],"minor_comments":[{"comment":"The reference to 'Fig. ??' in the qualitative-results paragraph is unresolved; it should point to Fig. 7.","section":"Section 4.2"},{"comment":"The caption 'Qualtitative Results' is a typo; it should read 'Qualitative Results'.","section":"Section 4.2"},{"comment":"The 'Random' row is not described in the text; please state how the random offset field T^k was sampled (distribution, scale) and whether the reported 24.58 dB is an average over multiple random draws.","section":"Table 3"},{"comment":"The average orientation vector in Eq. (5) is computed by averaging the per-pixel horizontal and vertical components without magnitude weighting, while Eq. (6) uses a magnitude-weighted mean for the global motion vector; please clarify this distinction in the text, as it may affect the orientation-similarity check used in ExCG.","section":"Section 3.2, Eq. (5)"},{"comment":"The phrase 'average gains of 2.24 dB on EVRB, 2.35 dB on REVD' is ambiguous; please specify whether the average is over all five backbones and clarify in the text.","section":"Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is well-written and the core idea is appealing, but the evaluation-protocol issue is load-bearing. If the adaptation frames are the same as the evaluation frames, the 'unseen domain' claim is not established; the authors must demonstrate gains under a disjoint train/test split. Additionally, tuning hyperparameters on REVD and then reporting REVD in the main tables is a circularity that needs to be addressed. I would be willing to reconsider after a revision that separates adaptation and evaluation data and reports variance across seeds."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read on arXiv:2608.08066 (EvBS). The core idea is real: use events to decouple motion from visual content, so you can re-blur a sharp patch with its own motion (intrinsic) and also warp a motion pattern from a different patch onto new content (extrinsic). Algorithm 1—the cross-patch motion transfer via event warping—is new relative to DADeblur, MetaDeblur, and ID-Blau, and the framework supports both event-based and frame-only backbones, which prior synthesis methods don't. The experiment set is thorough: five models, five benchmarks, consistent gains around 2–2.6 dB PSNR over baselines, and ablations (Table 3, Fig. 9) make a credible case for each component. The supplementary is honest about label noise, sensor misalignment, and cost.\n\nThe soft spot is not the mechanism; it's the evaluation. The Dual Source Extractor pulls content and motion sources from the same target frames used for evaluation, synthesizes training pairs from those exact frames, fine-tunes on them, and reports metrics on that same dataset with no scene- or frame-level split. That turns 'unseen testing dataset' into 'test-time adaptation on the target frames, evaluated on those frames'—a legitimate TTA setting, but one where part of the gain may come from the model learning the specific synthesis mapping rather than a generalizable domain shift. The REVD numbers are also optimistically biased because r, n, and orientation thresholds are tuned on REVD, which is in the main tables. No error bars, no significance tests, no code release.\n\nNone of this undermines the core idea; the mechanism is sound and the ablations support it. But the claim of improving unseen domains needs a cleaner protocol: adapt on one subset of the target data and evaluate on a held-out subset, or leave-one-scene-out, and report variance. I'd send this to peer review—the idea deserves referee time—but I'd make the evaluation protocol the major revision point.","headline":"A genuinely novel event-based mechanism for transferring motion across patches, with consistent empirical gains, but the evaluation protocol blurs the line between adaptation and memorization because the same frames serve as both training and test data.","tokens_in":17901,"tokens_out":4977,"would_cite":true,"duration_ms":47140,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"By using events to decouple motion from visual content, EvBS synthesizes diverse blurry training pairs that push pre-trained deblurring models to perform better on unseen real-world domains.","keywords":["event camera","blur synthesis","domain adaptation","motion deblurring","diffusion model","optical flow","event-guided deblurring","test-time adaptation"],"falsifier":"Construct a synthetic target domain with known ground-truth flow, for example scenes rendered with ESIM under scripted motion, and run EvBS's extrinsic transfer. If the flow of the synthesized blur does not match the intended motion source's flow beyond the orientation threshold, the decoupling claim fails. A simpler behavioral test is a scene with only uniform global camera motion, where every patch has the same flow and extrinsic transfer adds no diversity: if EvBS still improves over intrinsic-only fine-tuning there, the benefit must come from something other than motion diversity.","tokens_in":16855,"feed_emoji":"⚡","tokens_out":4238,"duration_ms":41801,"temperature":0.7,"pith_summary":"The paper claims that the reason pre-trained motion deblurring models fail on new real-world data is the domain gap, and that existing domain-adaptation schemes are handicapped because motion and visual content are entangled inside a single RGB exposure, so they can only re-blur a patch with its own motion. EvBS uses event streams, which record motion at microsecond resolution, to pull motion patterns apart from visual content. That lets it do two things: blur a sharp patch with its own event-derived motion (intrinsic), and inject motion from an unrelated blurry patch into a different sharp patch (extrinsic), creating many more sharp/blurry training pairs than natural video provides. Fine-tuning a pre-trained deblurrer on these pairs, with the original sharp patches as supervision, raises PSNR on unseen real-world event-based benchmarks by roughly 2.2 to 2.6 dB across both event-based and frame-only architectures. If the claim is right, this gives a plug-in recipe for adapting existing deblurring models to a new scene or camera without requiring any ground-truth sharp/blurry pairs.","feed_headline":"Events decouple motion from content to retrain deblurring models","feed_subtitle":"Mixes motions from one patch with sharp content from another, lifting PSNR on unseen domains by 2+ dB.","key_machinery":"The central object is the event-guided blur-conditioning pipeline. A Dual Source Extractor (DSE) locates pseudo-sharp content patches (low event count and low predicted blur) and coherent-motion blur patches (high blur score with direction-consistent flow), using a blur score map that combines an event count map with a retrained blur magnitude estimator. A Blur Conditioning Module (BCM) then builds conditions through two generators: the Intrinsic-blur Condition Generator (InCG), which forms conditions from the content patch's own event-derived flows plus a Flow-Enhanced Deviation Accumulation (FEDA) representation, and the Extrinsic-blur Condition Generator (ExCG), which transfers motion by filtering on angular difference between content and motion orientation, then warping events via Algorithm 1. These conditions drive a conditional diffusion reblurring model (a modified ID-Blau) to synthesize blurred images, whose original sharp content serves as supervision for fine-tuning the deblurrer.","core_discovery":"EvBS establishes that events supply a motion-content decoupling that RGB frames cannot, and that this decoupling directly unlocks combinatorial blur synthesis. The central discovery is that you can take a pseudo-sharp content source and a separate motion source from the same target domain and, by warping the content source's events along an offset field computed from the motion source's global flow minus the content source's local flow, produce a training pair whose blur follows the transferred motion. The authors verify the transfer qualitatively by showing the synthesized flow follows the motion source rather than the content source, and quantitatively by ablating random offsets, which drop PSNR below the no-adaptation baseline. Consequently, fine-tuning on these synthetic pairs improves both event-based models (EFNet, MAENet, EBFI) and frame-only models (NAFNet, MIMO-UNet), with the largest gains around 2.6 dB PSNR on the HighREV benchmarks.","pith_inferences":["A natural extension is that the same motion-content decoupling could serve domain adaptation for other motion-sensitive tasks, such as event-based optical flow or detection under motion blur, because it supplies dense flow and dynamics supervision without ground truth.","The orientation filter range of 15 to 30 degrees suggests a sweet spot between redundancy and realism; on data with mostly unidirectional camera motion the extrinsic pathway may add little diversity, so gains should be larger on scenes with mixed object and ego motion, which is testable by splitting target datasets by motion direction.","Because the pipeline relies on a patch being pseudo-sharp, it inherits a ceiling: if a target scene has no truly still regions, supervision quality drops; an adaptive threshold or multi-frame sharp fusion could push this ceiling higher.","The reported offline cost (ExCG at 428.8 seconds versus InCG at 61.5 seconds on EVRB) implies an InCG-only lightweight variant is practical, and a deployment could gate ExCG on whether the scene shows diverse motion directions."],"forward_implications":["Fine-tuning with EvBS pairs improves five pre-trained deblurring models on EVRB, REVD, and HighREV variants, with average PSNR gains of roughly 2.2 to 2.6 dB while also improving SSIM and LPIPS.","Extrinsic blur synthesis alone contributes about as much as intrinsic synthesis (28.14 vs 28.03 dB on REVD with MAENet), and combining both gives the best result (28.64 dB), showing that motion diversity is the driver of adaptation.","The method transfers to frame-only architectures even though it is designed around events, making it a general adaptation add-on rather than an event-specific one.","The FEDA representation beats both flow-only and plain deviation accumulation for blur synthesis quality and for downstream adaptation performance.","Random motion offsets during transfer drop PSNR below the no-adaptation baseline (24.58 vs 25.41 dB), confirming that motion-aware transfer, not arbitrary augmentation, is what helps."],"supporting_citations":[{"why":"E-RAFT supplies all event-based optical flow estimates used by InCG, ExCG, and the motion source selector.","marker":"[7]"},{"why":"He et al. provide the blur magnitude estimator, the adaptive patch-selection scheme, and the test-time-blurring baseline that EvBS extends and compares against.","marker":"[8]"},{"why":"ID-Blau is the conditional diffusion reblurring model that EvBS modifies and retrains with event-guided conditioning.","marker":"[41]"},{"why":"Deviation accumulation from Sun et al. is the base representation that FEDA extends by weighting with flow magnitude.","marker":"[36]"},{"why":"Meta Transferring for Deblurring is the synthesis-based domain-adaptation baseline that EvBS must beat, and its limitations motivate the extrinsic pathway.","marker":"[21]"},{"why":"The event camera sensor paper establishes the microsecond temporal resolution premise that makes motion-content decoupling possible.","marker":"[20]"},{"why":"The realistic blur synthesis pipeline is used to generate the HighREV-3, HighREV-7, and HighREV-11 benchmarks that EvBS evaluates on.","marker":"[31]"}],"fun_headline_variants":["EvBS: Events decouple motion and content for blur synthesis","Event-guided blur synthesis improves domain-adaptive deblurring","Move blur from one frame to another with event cameras","Transfer motion via events to retrain deblurring networks","Event-driven blur mixing boosts deblurring on unseen domains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline assumes that the detected content sources are genuinely sharp enough to serve as clean supervision and that the event stream is precisely aligned with those frames; the paper itself notes that residual blur in selected patches adds label noise (Sec 4.3) and that sensor misalignment degrades performance (Supplementary Sec E).","fun_headline_variants_meta":{"raw":{"variants":["EvBS: Events decouple motion and content for blur synthesis","Event-guided blur synthesis improves domain-adaptive deblurring","Move blur from one frame to another with event cameras","Transfer motion via events to retrain deblurring networks","Event-driven blur mixing boosts deblurring on unseen domains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00022,"raw_usage":{"total_tokens":1453,"prompt_tokens":958,"completion_tokens":495,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":411}},"tokens_in":574,"tokens_out":495,"duration_ms":5460,"temperature":1.0,"reasoning_tokens":411,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:27:21.260968+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a synthetic target domain with known ground-truth flow, for example scenes rendered with ESIM under scripted motion, and run EvBS's extrinsic transfer. If the flow of the synthesized blur does not match the intended motion source's flow beyond the orientation threshold, the decoupling claim fails. A simpler behavioral test is a scene with only uniform global camera motion, where every patch has the same flow and extrinsic transfer adds no diversity: if EvBS still improves over intrinsic-only fine-tuning there, the benefit must come from something other than motion diversity.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"E-RAFT supplies all event-based optical flow estimates used by InCG, ExCG, and the motion source selector."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"He et al. provide the blur magnitude estimator, the adaptive patch-selection scheme, and the test-time-blurring baseline that EvBS extends and compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ID-Blau is the conditional diffusion reblurring model that EvBS modifies and retrains with event-guided conditioning."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Deviation accumulation from Sun et al. is the base representation that FEDA extends by weighting with flow magnitude."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Meta Transferring for Deblurring is the synthesis-based domain-adaptation baseline that EvBS must beat, and its limitations motivate the extrinsic pathway."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The event camera sensor paper establishes the microsecond temporal resolution premise that makes motion-content decoupling possible."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The realistic blur synthesis pipeline is used to generate the HighREV-3, HighREV-7, and HighREV-11 benchmarks that EvBS evaluates on."}],"review_version":1}