DeblurSplat: SfM-Free, Event-Assisted 3D Deblurring

• DeblurSplat is a framework for deblurring 3D scenes that combines event streams and dense stereo priors to overcome motion blur challenges.
• It integrates a DUSt3R-based SfM-free initialization with Confidence Balanced Sampling to efficiently generate robust point clouds and precise 3D Gaussian optimization.
• The method decodes latent sharp images via event-based supervision, enabling accurate blur trajectory estimation and improved novel view synthesis under adverse conditions.

DeblurSplat is a Structure-from-Motion-free, event-assisted 3D Gaussian Splatting framework for robust deblurring and high-quality novel view synthesis from motion-blurred images. It directly uses blurred RGB images and synchronous event streams, combines a pretrained dense stereo module with event-stream-based latent sharp image decoding, and jointly optimizes 3D Gaussians with blur trajectories. In the source paper, it is presented as the first Structure-from-Motion-free deblurring 3D Gaussian Splatting method via event camera (Li et al., 23 Sep 2025).

1. Problem setting and formulation

DeblurSplat addresses motion blur in 3D Gaussian Splatting under conditions such as low light or long exposure, where a captured frame integrates many sharp observations during the exposure interval. In this setting, edges and textures become smeared, large regions exhibit low contrast, and multi-view reconstruction becomes ambiguous. Standard 3D Gaussian Splatting and NeRF pipelines, when trained directly on such inputs, tend to misestimate geometry, produce blurry or ghosted novel views, and fail under complex motion or low-texture conditions (Li et al., 23 Sep 2025).

A central motivation of the method is the fragility of Structure-from-Motion-based initialization under blur. Existing deblur-3D-GS and deblur-NeRF pipelines typically rely on COLMAP to detect and match local features, estimate camera poses, and triangulate an initial point cloud. Under motion blur, feature detection and matching become unreliable, camera poses become inaccurate, and these errors propagate into the initial point cloud and subsequent optimization. DeblurSplat replaces this step with an SfM-free initialization and supplements it with event supervision, exploiting the blur-free, high-temporal-resolution brightness changes recorded by an event camera (Li et al., 23 Sep 2025).

The method also assumes the standard motion-blur image formation model used throughout the deblurring literature: a blurred frame is a time average of latent sharp frames. In DeblurSplat, this idea is instantiated inside a Gaussian Splatting renderer by introducing latent sharp renderings across the exposure interval and optimizing them against both the blurred RGB input and event-derived latent supervision (Li et al., 23 Sep 2025).

2. SfM-free initialization with DUSt3R and Confidence Balanced Sampling

The first stage of DeblurSplat replaces classical SfM with DUSt3R, a pretrained dense stereo module. Applied to blurred RGB images, DUSt3R predicts dense pointmaps and pixel-wise confidence maps, and the pipeline then recovers approximate intrinsics, relative transformations, and a globally aligned dense point cloud. This design avoids the cumulative error transfer from inaccurate camera poses to the initial point cloud positions, which the paper identifies as a major failure mode of blur-sensitive SfM pipelines (Li et al., 23 Sep 2025).

DeblurSplat does not use all DUSt3R points directly. Instead, it introduces Confidence Balanced Sampling (CBS), which partitions the confidence range $[c_{\min}, c_{\max}]$ into $M$ bins,

$$B_m = [c_{\min} + (m-1)\Delta,\; c_{\min} + m\Delta), \qquad \Delta = \frac{c_{\max} - c_{\min}}{M},$$

defines subsets

$$\mathcal{S}_m = \{ \mathbf{p}_k \in \mathcal{P} \mid c_k \in B_m \},$$

and samples $\frac{L}{M}$ points from each subset with probability proportional to confidence. The sampled set is

$$\mathcal{P}_{\text{sampled}} = \bigcup_{m=1}^M \left\{ \mathbf{p}_k \mid \mathbf{p}_k \sim \mathbb{P}(\mathcal{S}_m),\; \mathbb{P}(\mathbf{p}_k) \propto c_k \right\}.$$

This produces an initialization that preserves dense, high-confidence regions while retaining lower-confidence points needed for global structure, rather than overconcentrating Gaussians in a few highly confident areas (Li et al., 23 Sep 2025).

The sampled points are used as the initial Gaussian means. The paper reports that this DUSt3R-based initialization is approximately $2.96\times$ faster than COLMAP even when COLMAP uses sharp images, and that CBS reduces the number of training steps needed relative to BAD-GS while preserving rendering efficiency above $210$ FPS (Li et al., 23 Sep 2025).

3. Event stream integration and latent sharp image decoding

The second pillar of DeblurSplat is event-based supervision. Events are generated asynchronously when the log-intensity change at a pixel exceeds a threshold $\Theta$, with polarity

$$p_{x,y,\tau} = \begin{cases} -1, & \log(\mathcal{I}_{x,y,\tau}) - \log(\mathcal{I}_{x,y,\tau-\Delta\tau}) < -\Theta \ +1, & \log(\mathcal{I}_{x,y,\tau}) - \log(\mathcal{I}_{x,y,\tau-\Delta\tau}) > \Theta . \end{cases}$$

During the exposure of a blurred image, the event stream is partitioned into $M$0 temporal bins $M$1 (Li et al., 23 Sep 2025).

DeblurSplat adopts the Event-based Double Integral formulation to decode latent sharp images from the blurred frame and its event stream. If $M$2 denotes the blurred image and $M$3 the latent sharp image at the start of exposure, then

$M$4

which yields

$M$5

and, for $M$6,

$M$7

Thus, each blurred frame is converted into a sequence of latent sharp images $M$8 spanning the exposure interval (Li et al., 23 Sep 2025).

These decoded latent images play two roles. First, they provide fine-grained temporal supervision for the Gaussian renderer. Second, they stabilize blur-trajectory optimization by providing supervision at multiple time slices rather than only on the time-averaged blurred observation. The paper emphasizes that this is more informative than using a single deblurred frame per exposure, because the event stream supplies full temporal structure throughout the interval (Li et al., 23 Sep 2025).

4. Gaussian Splatting model, latent trajectories, and Progressive Alignment

DeblurSplat retains the standard 3D Gaussian Splatting representation. Each Gaussian has a mean position $M$9, covariance $$B_m = [c_{\min} + (m-1)\Delta,\; c_{\min} + m\Delta), \qquad \Delta = \frac{c_{\max} - c_{\min}}{M},$$0, opacity $$B_m = [c_{\min} + (m-1)\Delta,\; c_{\min} + m\Delta), \qquad \Delta = \frac{c_{\max} - c_{\min}}{M},$$1, and color $$B_m = [c_{\min} + (m-1)\Delta,\; c_{\min} + m\Delta), \qquad \Delta = \frac{c_{\max} - c_{\min}}{M},$$2. The covariance is parameterized as

$$B_m = [c_{\min} + (m-1)\Delta,\; c_{\min} + m\Delta), \qquad \Delta = \frac{c_{\max} - c_{\min}}{M},$$3

with scale $$B_m = [c_{\min} + (m-1)\Delta,\; c_{\min} + m\Delta), \qquad \Delta = \frac{c_{\max} - c_{\min}}{M},$$4 and rotation $$B_m = [c_{\min} + (m-1)\Delta,\; c_{\min} + m\Delta), \qquad \Delta = \frac{c_{\max} - c_{\min}}{M},$$5 encoded by a quaternion. Under a camera pose $$B_m = [c_{\min} + (m-1)\Delta,\; c_{\min} + m\Delta), \qquad \Delta = \frac{c_{\max} - c_{\min}}{M},$$6, the projected 2D covariance is

$$B_m = [c_{\min} + (m-1)\Delta,\; c_{\min} + m\Delta), \qquad \Delta = \frac{c_{\max} - c_{\min}}{M},$$7

and the pixel color is obtained by front-to-back alpha compositing,

$$B_m = [c_{\min} + (m-1)\Delta,\; c_{\min} + m\Delta), \qquad \Delta = \frac{c_{\max} - c_{\min}}{M},$$8

with $$B_m = [c_{\min} + (m-1)\Delta,\; c_{\min} + m\Delta), \qquad \Delta = \frac{c_{\max} - c_{\min}}{M},$$9 and $$\mathcal{S}_m = \{ \mathbf{p}_k \in \mathcal{P} \mid c_k \in B_m \},$$0 (Li et al., 23 Sep 2025).

What changes is the temporal rendering model. For each blurred image $$\mathcal{S}_m = \{ \mathbf{p}_k \in \mathcal{P} \mid c_k \in B_m \},$$1, DeblurSplat introduces a pose trajectory during exposure. Let $$\mathcal{S}_m = \{ \mathbf{p}_k \in \mathcal{P} \mid c_k \in B_m \},$$2 be the DUSt3R initialization for that frame, and let $$\mathcal{S}_m = \{ \mathbf{p}_k \in \mathcal{P} \mid c_k \in B_m \},$$3 and $$\mathcal{S}_m = \{ \mathbf{p}_k \in \mathcal{P} \mid c_k \in B_m \},$$4 be trainable start- and end-of-exposure pose adjustments. The latent pose at exposure step $$\mathcal{S}_m = \{ \mathbf{p}_k \in \mathcal{P} \mid c_k \in B_m \},$$5 is

$$\mathcal{S}_m = \{ \mathbf{p}_k \in \mathcal{P} \mid c_k \in B_m \},$$6

At each $$\mathcal{S}_m = \{ \mathbf{p}_k \in \mathcal{P} \mid c_k \in B_m \},$$7, the current Gaussian scene renders a latent sharp image $$\mathcal{S}_m = \{ \mathbf{p}_k \in \mathcal{P} \mid c_k \in B_m \},$$8, and the synthetic blurred image is the average

$$\mathcal{S}_m = \{ \mathbf{p}_k \in \mathcal{P} \mid c_k \in B_m \},$$9

This is the coarse component of what the paper calls Progressive Alignment (Li et al., 23 Sep 2025).

The blur reconstruction loss is

$\frac{L}{M}$0

with $\frac{L}{M}$1. The fine component uses the EDI-decoded latent sharp images, but compares them in grayscale to reduce sensitivity to color mismatch:

$\frac{L}{M}$2

where $\frac{L}{M}$3 denotes grayscale conversion. The total objective is

$\frac{L}{M}$4

with $\frac{L}{M}$5 in the reported implementation (Li et al., 23 Sep 2025).

Optimization uses Adam for 20,000 iterations. The first 3,000 iterations use only $\frac{L}{M}$6 as a warm-up; after that, $\frac{L}{M}$7 is added. The implementation uses $\frac{L}{M}$8 latent frames per blurred image and event threshold $\frac{L}{M}$9. In the paper’s interpretation, the coarse loss aligns time-averaged blur, while the event-guided fine loss aligns latent sharp views, progressively correcting both geometry and blur trajectories (Li et al., 23 Sep 2025).

5. Position within the deblur-Gaussian-Splatting literature

DeblurSplat belongs to a broader family of methods that perform deblurring directly inside Gaussian Splatting rather than as a preprocessing step. Earlier 3DGS deblurring formulations include camera-motion blur modeling with Gaussian Densification Annealing in DeblurGS (Oh et al., 2024), covariance manipulation for defocus blur in "Deblurring 3D Gaussian Splatting" (Lee et al., 2024), event-assisted Gaussian-center correction with an Adaptive Deviation Estimator in EaDeblur-GS (Weng et al., 2024), and event-driven bundle-adjusted blur modeling in EBAD-Gaussian (Deng et al., 14 Apr 2025).

A distinct branch extends blur-aware Gaussian Splatting beyond static reconstruction. Deblur-SLAM formulates sub-frame camera trajectory estimation, loop closure, and global bundle adjustment in a blur-aware RGB SLAM pipeline (Girlanda et al., 16 Mar 2025); BARD-GS decomposes blur into camera motion blur and object motion blur for dynamic scenes (Lu et al., 20 Mar 2025); Deblur4DGS reconstructs a 4D Gaussian representation from blurry monocular video by turning continuous within-exposure motion estimation into an exposure-time estimation problem (Wu et al., 2024). Robust Gaussian Splatting also addresses motion blur, defocus blur, pose refinement, and color inconsistencies in a unified 3DGS formulation (Darmon et al., 2024).

Within that landscape, DeblurSplat is distinguished by two design choices stated explicitly in the paper: it is SfM-free, because it initializes from DUSt3R rather than classical SfM, and it is event-assisted, because it uses event-decoded latent sharp images as a fine-grained supervision signal during Gaussian and pose optimization (Li et al., 23 Sep 2025).

6. Experimental evaluation

The paper evaluates DeblurSplat on three categories of data: normal synthetic scenes from Deblur-NeRF, hard synthetic scenes rendered in Blender with complex geometry and large motion, and real-world recordings from a Color DAVIS346 event camera mounted on a UR-5E robotic arm with nonlinear motion. The real dataset includes five tasks and five blur levels with exposure times $$\mathcal{P}_{\text{sampled}} = \bigcup_{m=1}^M \left\{ \mathbf{p}_k \mid \mathbf{p}_k \sim \mathbb{P}(\mathcal{S}_m),\; \mathbb{P}(\mathbf{p}_k) \propto c_k \right\}.$$0, $$\mathcal{P}_{\text{sampled}} = \bigcup_{m=1}^M \left\{ \mathbf{p}_k \mid \mathbf{p}_k \sim \mathbb{P}(\mathcal{S}_m),\; \mathbb{P}(\mathbf{p}_k) \propto c_k \right\}.$$1, $$\mathcal{P}_{\text{sampled}} = \bigcup_{m=1}^M \left\{ \mathbf{p}_k \mid \mathbf{p}_k \sim \mathbb{P}(\mathcal{S}_m),\; \mathbb{P}(\mathbf{p}_k) \propto c_k \right\}.$$2, $$\mathcal{P}_{\text{sampled}} = \bigcup_{m=1}^M \left\{ \mathbf{p}_k \mid \mathbf{p}_k \sim \mathbb{P}(\mathcal{S}_m),\; \mathbb{P}(\mathbf{p}_k) \propto c_k \right\}.$$3, and $$\mathcal{P}_{\text{sampled}} = \bigcup_{m=1}^M \left\{ \mathbf{p}_k \mid \mathbf{p}_k \sim \mathbb{P}(\mathcal{S}_m),\; \mathbb{P}(\mathbf{p}_k) \propto c_k \right\}.$$4, with the shortest exposure used as near-sharp reference (Li et al., 23 Sep 2025).

On hard synthetic scenes, DeblurSplat reports average PSNR $$\mathcal{P}_{\text{sampled}} = \bigcup_{m=1}^M \left\{ \mathbf{p}_k \mid \mathbf{p}_k \sim \mathbb{P}(\mathcal{S}_m),\; \mathbb{P}(\mathbf{p}_k) \propto c_k \right\}.$$5, SSIM $$\mathcal{P}_{\text{sampled}} = \bigcup_{m=1}^M \left\{ \mathbf{p}_k \mid \mathbf{p}_k \sim \mathbb{P}(\mathcal{S}_m),\; \mathbb{P}(\mathbf{p}_k) \propto c_k \right\}.$$6, and LPIPS $$\mathcal{P}_{\text{sampled}} = \bigcup_{m=1}^M \left\{ \mathbf{p}_k \mid \mathbf{p}_k \sim \mathbb{P}(\mathcal{S}_m),\; \mathbb{P}(\mathbf{p}_k) \propto c_k \right\}.$$7, compared with EDNeRF at PSNR $$\mathcal{P}_{\text{sampled}} = \bigcup_{m=1}^M \left\{ \mathbf{p}_k \mid \mathbf{p}_k \sim \mathbb{P}(\mathcal{S}_m),\; \mathbb{P}(\mathbf{p}_k) \propto c_k \right\}.$$8, SSIM $$\mathcal{P}_{\text{sampled}} = \bigcup_{m=1}^M \left\{ \mathbf{p}_k \mid \mathbf{p}_k \sim \mathbb{P}(\mathcal{S}_m),\; \mathbb{P}(\mathbf{p}_k) \propto c_k \right\}.$$9, and LPIPS $2.96\times$0. On normal synthetic scenes, DeblurSplat reports PSNR $2.96\times$1, SSIM $2.96\times$2, and LPIPS $2.96\times$3. On real scenes, the reported averages are PSNR $2.96\times$4, SSIM $2.96\times$5, and LPIPS $2.96\times$6, compared with EDNeRF at PSNR $2.96\times$7 and LPIPS $2.96\times$80.7129$$(<a href="/papers/2509.18898" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">Li et al., 23 Sep 2025</a>).</p> <p>The paper also reports blur-level analysis showing that DeblurSplat consistently outperforms B-3DGS, E²GS, BAD-GS, and EDNeRF across blur levels$$2.96\times$9 through $210$0, including the highest reported exposure setting. In the ablations, CBS alone improves PSNR over COLMAP-initialized B-3DGS by about $210$1 to $210$2 dB, Progressive Alignment yields a large improvement over coarse-only training, balanced sampling performs slightly better than spatial, random, and centered sampling, and sampling too many initial points, such as $210$3k or $210$4k, degrades convergence relative to $210$5k (Li et al., 23 Sep 2025).

The reported runtime characteristics are also part of the method’s positioning. Initialization is about $210$6 faster than COLMAP, CBS reduces the number of training steps needed relative to BAD-GS, and all configurations are reported to retain rendering above $210$7 FPS. In the paper’s framing, this combines high-fidelity novel view synthesis with significant rendering efficiency compared to prior deblur 3D-GS systems (Li et al., 23 Sep 2025).

7. Limitations and future directions

The paper identifies several limitations. First, the reliability of DUSt3R confidence maps under severe blur is not fully characterized, even though CBS assumes that the confidence values remain meaningful for balancing point-cloud sampling. Second, the explicit blur trajectory model is primarily camera-centric and is not designed for more complex scene dynamics, multiple moving objects, or rolling-shutter effects. Third, EDI and event binning assume event streams of sufficient quality; noisy or sparse events can affect latent image decoding. Fourth, scaling the combined DUSt3R, event, and blur-trajectory pipeline to very large scenes or thousands of images requires further engineering (Li et al., 23 Sep 2025).

The future directions discussed in the paper follow directly from those limitations. They include a more rigorous quality assessment for pointmaps under heavy blur, extending the framework to more complex dynamics and rolling shutter, incorporating more advanced event processing or event-to-video priors, and improving scalability to larger datasets. A plausible implication is that DeblurSplat’s main contribution is not only a specific pipeline, but also a template for combining dense stereo priors, event-derived latent supervision, and blur-aware Gaussian optimization in a single 3D reconstruction system (Li et al., 23 Sep 2025).