acceptodds
Withdrawn from ICLR 2027

SMART-GS: Blur-Guided Gaussian Splatting reconstruction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

A blurred image does not tell 3D Gaussian Splatting how to learn the sharp details, while a deblurring network could not add enough details that are consistent across views. Both problems are strongly related and a joint optimization by no means straightforward. We propose SMART-GS, an offline deblurring 3DGS method in which image blur guides the optimization process without assuming a particular type of blur. We use the amount of blur in the input image to define how much each observed frame and a corresponding deblurred image should supervise the optimization method. The blurriness of the current rendered image then measures how many sharp details have been recovered which we use to update the supervision loss for recovering more details in the blurred image. We also use the blurriness of both images to guide 3DGS densification so the optimizer learns both what to trust and where to grow in a complex input video sequence. One set of sharp 3D Gaussian representations explains captured and deblurring observations through ordered, learned blur formation. At inference, only this sharp representation remains. SMART-GS reaches state-of-the-art deblurring performance on five real- world benchmarks: motion blur, defocus blur, extreme motion blur, TUM-RGBD, and the ScanNet dataset. We achieve 27.74 dB PSNR higher than the state-of-the- art methods on the extreme motion blur sequences. We also study how supervision and capacity affect the optimization of large indoor scenes.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Deblurring, Novel View Synthesis, Adaptive Supervision, Learned Density Control

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_wpDcOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SMART-GS which leverages single-image deblurring priors and rendered recovery to dynamically guide blur-adaptive supervision and reinforcement learning-based Gaussian densification

Critical strengths

The paper uses rendered detail recovery to balance 2D deblurring priors against blurry inputs and guide RL-based Gaussian densification, recovering sharp details without introducing view-inconsistent 3D artifacts.

Critical weaknesses
  1. Figure 2 fails to clearly convey the technical pipeline. It should illustrate concrete methodological details rather than remain at a conceptual level.
  2. Regarding the EVSSM deblurring model used in the paper, was it retrained or fine-tuned on target datasets? In addition, the comparison of the total reconstruction time should be provided.
Questions for authors

I have to say that the obvious signs of AI assistance in Figures 1 and 2 hurt my impression of the work, although the authors have claimed it in AI USE STATEMENT.

Reviewer_ZPjqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SMART-GS for reconstructing a sharp 3D Gaussian scene from multi-view images containing motion blur, defocus blur, or mixed blur.

For each input view, a pretrained single-image deblurring model first produces a deblurred image, which is treated as a prior. SMART-GS then estimates how much this prior should be trusted relative to the original captured image. During optimization, the confidence is updated according to whether the current 3D render has recovered image details that are supported by the input. A shared blur proposal network is used to explain both the captured image and the deblurred prior.

The same blur/recovery information is also incorporated into a reinforcement-learning density controller derived from LeGS, which chooses keep, clone, split, and prune operations for the Gaussians.

Critical strengths
  1. The problem is practically relevant. Real multi-view captures often contain heterogeneous motion and defocus blur, and directly fitting 3DGS to such observations can produce blurry or inconsistent reconstructions.

  2. The most interesting part of the method is the adaptive supervision between the original blurry observation and the deblurred prior. Rather than treating the output of the 2D deblurring network as ground truth, the method estimates an initial confidence and gradually changes the supervision according to the recovery state of the 3D reconstruction.

  3. Extending learned Gaussian density control with blur and recovery information is also a reasonable idea. The M4-to-M5 comparison suggests that the blur-aware extension of LeGS can improve reconstruction quality.

  4. The paper evaluates several types of real captured data and includes recent baselines such as Unblur-SLAM (CVPR 2026) and PRISM3D (ECCV 2026).

Critical weaknesses

1. The main quantitative protocol does not validate the claimed novel-view / cross-view reconstruction quality.

All supplied source frames, including the frames used for the reported metrics, participate in scene optimization. The paper itself describes this as an observed-camera reconstruction protocol rather than a camera-disjoint novel-view benchmark.

This protocol is useful for measuring how well the method restores the observed cameras, but the paper motivates the method partly by arguing that independent 2D deblurring cannot guarantee cross-view consistency or geometrically correct novel views. Those claims are not directly validated by the current quantitative evaluation.

A held-out-camera evaluation would therefore be important. If geometry ground truth is available for the indoor datasets, geometry metrics would also help determine whether sharper appearance corresponds to a better 3D reconstruction.

2. The current ablation does not cleanly isolate the proposed contributions.

The cumulative ablation shows that M0-M3, which introduce the adaptive-supervision components, provide only small changes and sometimes reduce PSNR or SSIM. The largest jump occurs when the existing LeGS density controller is introduced in M4.

The comparison between M4 and M5 does support the blur-aware extension, but M5 also uses substantially more Gaussians: approximately 10.18M vs. 2.86M on ScanNet and 10.78M vs. 1.93M on TUM. Thus, the current experiment does not establish whether the improvement comes from better blur-aware density allocation or simply from a much larger representation.

A fixed-Gaussian-budget comparison between LeGS and the proposed blur-aware policy is needed. It would also be useful to separately ablate the additional blur-aware state, delayed reward, and structural checks.

A simple deblur-first baseline is also missing: use exactly the same pretrained deblurring model to process all input views, and then reconstruct the scene using the same underlying 3DGS optimizer/density controller. This would show how much is gained by the proposed adaptive confidence mechanism rather than simply by providing deblurred images.

3. The novelty should be positioned more narrowly, and several closely related works are missing.

Reconstructing sharp 3DGS from blurry multi-view images is not new. Unified handling of motion and defocus blur is also not new, and neither are 2D deblurring priors, blur-kernel estimation, or RL-based Gaussian density control individually.

The main novelty of SMART-GS appears to be more specific:
(1) dynamically adjusting the trust between the captured blurry observation and a deblurred image prior according to reconstruction progress; and
(2) injecting the same blur/recovery information into a LeGS-style density controller.

The related-work discussion should therefore include at least the following closely related methods:

  • Li Liu, Jing Duan, Xiaodong Fu, Wei Peng, and Lijun Liu. "Unified 3D Gaussian Splatting for Motion and Defocus Blur Reconstruction." Visual Informatics, 9(4):100270, 2025.
    This work already studies a unified 3DGS framework for motion and defocus blur reconstruction.

  • Qida Cao, Jiajun Ding, Zhenyang Liu, Zhenzhong Kuang, Yijie Shao, and Yilan Shen. "VC-GS: View-Consistent Deblurring Gaussian Splatting via Alternating Branch Optimization." The Visual Computer, 41(10):7303-7317, 2025.
    This work is particularly relevant because it also starts from 2D-deblurred multi-view inputs and explicitly addresses the inconsistency introduced by independent image deblurring.

The paper should clarify its novelty relative to these works rather than presenting unified motion/defocus handling itself as a primary contribution.

4. The external deblurring prior is insufficiently documented.

The manuscript cites Kong et al. (2025) for the single-image deblurring network, but the exact checkpoint and its training data are not specified.

This matters because the cited deblurring model is primarily evaluated on standard motion/real-world deblurring datasets, while SMART-GS also relies on its outputs for defocus-blurred images. Since the prior is central to the proposed supervision mechanism, the paper should report which checkpoint is used and how well the prior itself performs on the motion and defocus subsets.

The initial prior confidence is also based on several hand-designed cues and constants (color consistency, edge agreement, clipping, and Laplacian detail gain). The paper should show that this confidence formulation is useful compared with simpler alternatives such as fixed confidence or a simpler image-quality score.

Questions for authors
  1. Can the authors report quantitative results on held-out cameras?
    The current metrics are computed at cameras whose observations were used during scene optimization. Since cross-view consistency and novel-view reconstruction are important motivations of the paper, a camera-disjoint evaluation would substantially strengthen the claims.

  2. How much of the improvement comes from the deblurred prior itself?
    Please compare against a simple baseline that applies exactly the same pretrained deblurring model to all input views and reconstructs the scene using the same underlying optimizer and density controller, without the proposed adaptive confidence mechanism.

  3. Does the blur-aware density controller still outperform LeGS at the same Gaussian budget?
    M5 uses about 3.56x as many Gaussians as M4 on ScanNet and 5.57x as many on TUM. A fixed-capacity comparison is necessary to separate better density allocation from increased representation size.

  4. Can the adaptive-supervision components be evaluated more directly?
    The current M0-M3 cumulative results provide only small and non-monotonic changes. Please report simpler controls such as fixed prior/input weights, initial confidence without recovery feedback, and recovery feedback without the hand-designed initial confidence.

  5. What exact single-image deblurring checkpoint is used?
    Please specify its training data and report the quality of the resulting deblurred priors separately on motion-blur and defocus-blur subsets. In particular, how reliable is this prior on defocus blur?

  6. Can the authors clarify the supplementary-video protocol and computational cost?
    Please report final Gaussian count, training time, peak memory, and inference FPS under the same hardware. For video comparisons, please also clarify whether all methods use the same camera trajectory, number of rendered frames, and playback frame rate.

Reviewer_xPeROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SMART-GS for Gaussian splatting reconstruction from blurred imagery. The authors introduce 1. adaptove supervision that determines how much deblurred image should supervise the optimization method as reconstruction progresses. 2. the blurriness of the current rendered image, edge strength using laplacian responses, measures how many sharp details have been recovered to update the supervision loss for recovering more details in the blurred image. 3. a blur-guided reinforcement learning (RL) growing strategy.

Critical strengths
  1. The motivation of leveraging blurry and deblurred inputs and blur-aware RL grow strategy is intuitive for the deblur reconstructions.
  2. Using the recovery signals to guide optimization and Gaussian allocation is a reasonable system design.
  3. The proposed method achieves strong results on motion and defocus blur datasets with the observed views setting.
Critical weaknesses
  1. As declared in "reproducibility statement" by the authors, the results were conducted with the observed views protocal instead of novel view synthesis, which is essential for the broader 3D claims.
  2. The 57.09 dB PSNR, 1.00 SSIM after rounding, and 0.001 LPIPS on ExBluRF require a detailed protocol audit, as its extreme-blur results are also higher than the blur datasets.
  3. In ablation Table, M1-M3 brings slignt difference because of no Gaussian grow. The order should be reversed: first add the LeGS Gaussian-grow policy, then add each component to show each effectiveness, alongside with a qualitative comparison.
  4. What is the computation budget during training and inference?
  5. The paper lacks a limitation section for readers to understand the practical boundaries of the proposed method.
Questions for authors

Please see comments above.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.