{"id":"955fd871-de65-4255-aa2a-9802a6b22475","arxiv_id":"2509.18898","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A pose-free deblurring 3D Gaussian Splatting pipeline using DUSt3R point clouds, confidence-balanced sampling, and event-decoded latent image supervision.","lead":"DeblurSplat is a new 3D scene reconstruction method that removes motion blur by replacing the usual SfM initialization with a dense stereo module (DUSt3R) and feeding event-camera data into the optimization. On synthetic and real blurry datasets it reports higher novel-view quality and faster initialization than existing deblurring 3D Gaussian Splatting approaches.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim rests on incomplete baselines: closest event-based 3DGS deblurring method (EVA-Gaussians) is cited but not compared, and real-world benchmarks initialize all methods via the proposed DUSt3R pipeline.","rationale":"I read the paper as an empirical claim that DeblurSplat outperforms prior SOTA deblurring reconstruction methods. The most load-bearing condition for that claim is that the comparison set contains the strongest prior methods. It does not: EVA-Gaussians (Ref. [12]) is the direct predecessor in event-assisted 3DGS deblurring and is cited in Section II-A but never appears in any comparison table. The real-world comparison further confounds the SfM-free contribution because Appendix D initializes all methods with the proposed DUSt3R pipeline, so the real-scene +3.36 dB gain does not test the paper's claimed advantage over SfM initialization. I do not dismiss the reader's blur-model concern, but it is (a) shared with the EDI/BAD-GS line of work, and (b) partly self-consistent on the synthetic data, which was generated by linear SE(3) interpolation and RGB blending. The missing baseline is more decisive: if EVA-Gaussians were run under the same initialization and achieved comparable numbers, the primary SOTA claim would fall regardless of the blur-model assumption. Hence I recommend keeping the CONDITIONAL verdict, with the condition being this baseline experiment.","tokens_in":21198,"tokens_out":13611,"duration_ms":113157,"concrete_test":"Reproduce EVA-Gaussians [12] and run it on the paper's hard synthetic and real-world benchmarks (Tables I, II, VIII) under two initialization protocols: (1) its default COLMAP/SfM initialization, and (2) the DUSt3R initialization used for all baselines in Appendix D. Use identical train/test splits, evaluation metrics (PSNR/SSIM/LPIPS), and blur levels. If EVA-Gaussians achieves PSNR within 0.5 dB of DeblurSplat or higher under either initialization, the 'outperforms existing SOTAs' claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim 'outperforms existing SOTAs' (Tables I and II) is not yet established because Section IV-C's baseline list omits EVA-Gaussians [12], the closest prior work: event-stream-assisted 3D Gaussian Splatting from blurry images, which the paper itself discusses in Section II-A. Without this comparison, the reported margins (+2.74 dB hard synthetic, +3.36 dB real) may reflect an incomplete competitor set rather than a genuine advance. In addition, Appendix D states that for the real-world data COLMAP fails and 'we initialize these scenes by our method', so all methods are initialized with the proposed DUSt3R-derived point clouds; this eliminates the SfM-free-vs-SfM contrast and hands the proposed initialization to competitors without re-tuning them. Thus the empirical support for the headline SOTA claim is conditional on a single missing baseline. A concrete check: evaluate EVA-Gaussians on the same splits under both its native COLMAP initialization and the DUSt3R initialization used for other baselines. If it matches or exceeds DeblurSplat under either initialization, the SOTA claim fails; if not, the concern is resolved.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"DeblurSplat proposes a deblurring 3D Gaussian Splatting framework that replaces COLMAP initialization with DUSt3R dense-stereo point clouds and adds an event-based Progressive Alignment module. Blurred images are decoupled into latent sharp frames via EDI; rendered latent frames are aligned in grayscale with event-derived latent images, while a blurred reconstruction is formed by averaging latent renders. The camera trajectory is parameterized by SE(3) interpolation and optimized jointly with the Gaussians. Experiments on synthetic (normal/hard) and real-world datasets report consistent gains over B-NeRF, B-3DGS, UFP-GS, EFN-GS, E2GS, BAD-NeRF, BAD-GS, and EDNeRF.","tokens_in":21482,"tokens_out":6049,"duration_ms":45331,"significance":"If the empirical claims hold, the paper makes a useful contribution: dense-stereo initialization is a plausible remedy for COLMAP failure under blur, and event-based latent-image alignment gives a strong deblurring signal. The paper also introduces new challenging benchmarks (7 Blender hard scenes and 5 real scenes with multiple blur levels) and includes useful ablations of sampling strategy, interval count, event-loss weight, and pose accuracy. However, the headline SOTA claim is currently conditional on a missing closest competitor and on a real-world protocol in which all baselines are initialized by the proposed DUSt3R pipeline. The contribution is therefore promising but needs stronger empirical grounding before publication.","major_comments":[{"comment":"The claim that the method does not calculate camera poses is contradicted by the paper itself. Appendix A-B derives focal length, pairwise RANSAC+PnP poses, and globally aligned poses via Eq. (14); Eq. (5) parameterizes and optimizes latent poses; Table XII reports ATE. This is not SfM-free in the usual sense. Please revise the wording and clearly state that DUSt3R provides an initial pose estimate that is subsequently refined.","section":"Abstract and Section III-B / Appendix A-B"},{"comment":"EVA-Gaussians [12] is described as event-stream-assisted 3D Gaussian splatting from blurry images but is absent from the baseline list and all tables. Because it is the closest prior work, the SOTA claim in Tables I-II is not yet established. Add EVA-Gaussians under its native COLMAP initialization and under the DUSt3R-initialized protocol; report per-scene metrics. If it matches or exceeds DeblurSplat under either initialization, the SOTA claim is not supported.","section":"Section IV-C and Section II-A"},{"comment":"The real-world protocol weakens the comparison. Appendix D states COLMAP fails and 'we initialize these scenes by our method,' so every baseline is initialized from the proposed DUSt3R point cloud (1k random subsample) without re-tuning. In the hard synthetic scenes, by contrast, baselines receive sharp-image COLMAP initialization. This asymmetry makes it hard to attribute the +3.36 dB real and +2.74 dB hard gains to the proposed modules rather than to the initialization handout. Provide COLMAP-initialized baselines where possible, or explicitly report all methods under a common initialization and add an initialization-source sensitivity study.","section":"Section IV-B/IV-D and Appendix D"},{"comment":"The event-guided latent images are obtained by solving Eq. (4) from the same blurred input and events under EDI assumptions, so L_e is not supervision from independent sharp ground truth. Any error in the uniform-blur model, constant threshold Θ=0.27, bin count u, or the linear SE(3) interpolation in Eq. (5) is systematically reinforced by the alignment loss. Please add sensitivity analyses over Θ and u, and test with non-linear/interpolated ground-truth trajectories on synthetic scenes (where high-speed sharp frames are available) to quantify the bias.","section":"Section III-B, Eqs. (2)-(5)"}],"minor_comments":[{"comment":"Notation is inconsistent: Section III-A uses M intervals and L points, while Section IV-A says 'm=40 and Q_m=125'; also 'internals' should be 'intervals'.","section":"Section IV-A"},{"comment":"The exposure-time definition t_exp = t_start - t_end has the wrong sign; it should be t_end - t_start.","section":"Appendix A"},{"comment":"References [49] and [50] are duplicate entries for the same paper.","section":"References"},{"comment":"The caption mentions a proportion β that is never defined in the main text.","section":"Figure 2"},{"comment":"EDNeRF is cited as [29] in Section IV-C but as [30] in Section IV-D; please resolve the reference numbering.","section":"Section IV-D"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the venue. The main empirical risk is the omission of EVA-Gaussians, the closest event-based 3DGS deblurring work, and the real-world initialization protocol that gives every baseline the proposed DUSt3R point cloud. The authors should also consider whether the DUSt3R pose computation makes the 'SfM-free' framing defensible; if not, the novelty claims should be adjusted accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read.\n\nDeblurSplat is a reasonable engineering contribution, not a conceptual leap. The main idea is simple: use DUSt3R instead of COLMAP for initialization in deblurring 3D-GS, sample points by confidence balanced across bins, and use EDI-decoded latent sharp images (guided by events) as an auxiliary grayscale supervision while also refining latent poses. The confidence-balanced sampling is a nice practical detail, and the empirical motivation in Fig. 2 is worth having.\n\nWhat the paper does well: the experiments are extensive for a paper of this type. They build hard synthetic scenes with large camera motion, capture a real event dataset with multiple blur levels, and report consistent improvements across almost every scene and metric. The ablations show each component helps. And the efficiency gain over COLMAP (nearly 3x) is real and useful.\n\nNow the soft spots, in order of seriousness.\n\nFirst, the missing EVA-Gaussians comparison. The paper cites EVA-Gaussians as closely related event-assisted blurry 3D-GS work, but never compares against it. That is a real gap. The strong margins over BAD-GS and EDNeRF are nice, but the closest method could plausibly close or beat them. The stress-test note got this right. This is fixable—run EVA-Gaussians under both its native COLMAP init and the DUSt3R init—but without it the 'SOTA' claim is provisional.\n\nSecond, the real-world evaluation is rigged in their favor. Appendix D says COLMAP fails on their real scenes, so they initialize every method with DUSt3R point clouds. Baselines are not tuned for that, and the proposed method obviously is. It also undermines the 'SfM-free' contrast on the real data. On hard synthetic scenes they give baselines COLMAP from sharp images, which is actually generous to baselines, so those margins are more believable.\n\nThird, the abstract says 'without calculating camera poses,' but the method estimates poses from DUSt3R and refines them. That's an overstatement, and a reviewer will catch it.\n\nFourth, the EDI-derived latent sharp images are not independent ground truth. They come from the same blurred inputs, so if EDI's assumptions (fixed threshold, linear pose interpolation) are wrong, the supervision reinforces the bias. The grayscale alignment limit helps, and events carry independent temporal info, so it's not fatal, but it's under-tested—no validation against real sharp latent frames.\n\nNo code, no data, no error bars. For a paper with custom datasets and SOTA claims, that's a reproducibility problem.\n\nOverall: the core idea is sound and the empirical package is above average. The missing baseline and real-init confound are the load-bearing weaknesses, and both are addressable. This deserves peer review, not desk rejection, but I'd want the revision to add EVA-Gaussians, fix the real-world init protocol, and release code. A serious referee could turn this into a solid paper.\n\nI'd say reading group: yes, if you work on event-based 3DGS or deblurring. I'd cite it as an example of DUSt3R-based initialization, once the comparison issue is settled.\n\nRecommendation: send to review.","headline":"DeblurSplat is a solid engineering combination of DUSt3R and event-based deblurring for 3D-GS with strong reported results, but the SOTA claim is undercut by a missing nearest-baseline comparison and a real-world evaluation that gives every method the proposed initialization.","tokens_in":22021,"tokens_out":5375,"would_cite":true,"duration_ms":38973,"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":"DeblurSplat claims to be the first SfM-free deblurring 3D Gaussian Splatting method, replacing COLMAP with DUSt3R point-cloud initialization and adding event-camera latent-image supervision to achieve state-of-the-art novel-view quality on","keywords":["3D Gaussian Splatting","motion deblurring","event camera","SfM-free reconstruction","DUSt3R","novel view synthesis","EDI decoupling","bundle adjustment"],"falsifier":"Record a scene with a robotic arm executing a deliberately non-uniform trajectory during a long exposure while an event camera logs the true motion, and compare DeblurSplat's reconstructed latent frames against ground-truth high-speed sharp frames: if the recovered latent frames deviate exactly where the trajectory departs from linear SE(3) interpolation, the blur model, not the optimization, is the limiting factor.","tokens_in":21079,"feed_emoji":"🎥","tokens_out":6049,"duration_ms":47864,"temperature":0.7,"pith_summary":"The paper proposes DeblurSplat, the first Structure-from-Motion-free framework for deblurring 3D Gaussian Splatting. It argues that motion-blurred images can be reconstructed reliably by replacing SfM initialization with a pretrained dense stereo module that predicts point clouds and confidence scores directly, and by using event streams to recover latent sharp frames that supervise optimization. The central claim is that this combination avoids cumulative pose errors from blur and yields state-of-the-art novel-view quality and rendering speed. A sympathetic reader would care because real-world captures under low light or long exposure routinely suffer motion blur, and existing SfM-based deblurring pipelines degrade or fail in those settings.","feed_headline":"No pose solver needed: deblurred 3D-GS beats SOTA by 2.74 dB","feed_subtitle":"Dense stereo plus event-camera supervision turns blurry captures into sharp novel views with faster initialization.","key_machinery":"The central mechanism is the pairing of DUSt3R's dense-stereo point clouds (with confidence scores) with event-decoupled latent sharp images. DUSt3R provides blur-robust initialization without pose estimation; Confidence Balanced Sampling converts its confidence map into Gaussian primitives; EDI decouples each blurred frame into u latent sharp images; Progressive Alignment optimizes two endpoint SE(3) pose corrections per view (Eq. 5) and renders the u latent frames, imposing a grayscale alignment loss (Eq. 6) against the EDI-decoded frames and a blur-formation loss (Eq. 7) averaging the renders against the blurred frame. The event stream is the fine-grained supervision that resolves the amb","core_discovery":"DeblurSplat replaces the fragile SfM initialization stage of deblurring 3D-GS with DUSt3R, a pretrained dense stereo network that outputs, from blurred images directly, a set of 3D points and per-point confidence scores without estimating camera poses as an intermediate. The paper introduces Confidence Balanced Sampling to keep low-confidence but spatially informative points, and Progressive Alignment: each blurred image is decoupled via the Event-based Double Integral (EDI) into u latent sharp frames, while the camera trajectory during exposure is parameterized as a linear interpolation between two endpoint SE(3) corrections. The Gaussians render these latent frames; a grayscale alignment l","pith_inferences":["An editor's testable extension: the same DUSt3R-plus-event pipeline might extend to rolling-shutter distortion or other spatially varying degradations, since the event stream carries continuous motion information (the paper itself suggests rolling-shutter as future work).","The EDI decoupling assumes a fixed contrast threshold (Theta = 0.27); making Theta learnable per scene could improve robustness on real sensors where the threshold drifts, and it is a direct experiment to run.","If DUSt3R's confidence is unreliable on blurry inputs (the paper's own limitation), an alternative is to fuse multiple stereo hypotheses or to refine confidence during Gaussian training; this could be tested on synthetic scenes with known ground-truth geometry.","The grayscale alignment loss implicitly assumes latent images share edge structure with rendered ones; adding a color-aware or perceptual term might help on scenes with strong color fringes, though it could also introduce new failure modes."],"forward_implications":["If the method works as claimed, motion-blurred captures can be reconstructed without running COLMAP, removing a common failure point and cutting initialization time by roughly 3 times.","The event-guided Progressive Alignment provides a template for using continuous event streams to correct latent poses, not just deblur pixels.","The confidence-balanced sampling principle suggests that discarding low-confidence points from learned stereo is suboptimal; balanced sampling preserves geometry.","The method claims real-time rendering (above 210 FPS) after training, so deblurring 3D reconstruction could be used interactively.","Consistent gains on increasingly blurry real scenes suggest the approach degrades gracefully as blur worsens."],"fun_headline_variants":["Deblur 3D-GS without SfM using event camera","Event-based deblurring for 3D Gaussian Splatting, SfM-free","No pose estimation: DUSt3R and events deblur 3D-GS","Skip SfM, use events: sharper 3D-GS from blur","Robust deblurring for 3D-GS: SfM-free with events"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that each blurred image is exactly the average of u latent sharp frames related by an exponential event response with a constant threshold (Theta = 0.27) and a camera trajectory that is a linear SE(3) interpolation between two endpoint corrections; if real exposure motion is non-uniform or the EDI model is inaccurate, the latent sharp images used for supervision are systematically biased.","fun_headline_variants_meta":{"raw":{"variants":["Deblur 3D-GS without SfM using event camera","Event-based deblurring for 3D Gaussian Splatting, SfM-free","No pose estimation: DUSt3R and events deblur 3D-GS","Skip SfM, use events: sharper 3D-GS from blur","Robust deblurring for 3D-GS: SfM-free with events"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000323,"raw_usage":{"total_tokens":1639,"prompt_tokens":723,"completion_tokens":916,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":467,"completion_tokens_details":{"reasoning_tokens":803}},"tokens_in":467,"tokens_out":916,"duration_ms":7075,"temperature":1.0,"reasoning_tokens":803,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T15:39:35.357691+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record a scene with a robotic arm executing a deliberately non-uniform trajectory during a long exposure while an event camera logs the true motion, and compare DeblurSplat's reconstructed latent frames against ground-truth high-speed sharp frames: if the recovered latent frames deviate exactly where the trajectory departs from linear SE(3) interpolation, the blur model, not the optimization, is the limiting factor.","supporting_citations":[],"review_version":1}