Paper deep dive
ClearGS: Reliability-Aware Gaussian Splatting from Handheld Videos
Xuanzhi Liu, Xinyi Wu, Hang Pan, Wensi Huang, Zhenyao Wu, Ruize Han, Song Wang
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:We present ClearGS for 3D Gaussian Splatting (3DGS) from handheld videos with uneven viewpoint coverage and mixed frame quality. Rather than selecting frames with binary decisions, ClearGS uses Reliability-aware View Allocation (RVA) to assign graded raw-supervision weights based on appearance reliability, degradation risk, and geometric utility, while weakly reactivating useful suppressed frames to maintain trajectory coverage. Since weighting cannot restore details lost to blur or distortion, ClearGS further introduces Render-Guided In-Video Restoration (RIVR). The current 3DGS render provides a pose-aligned structural candidate, a frozen no-reference restoration expert restores the corresponding raw video observation without any clean reference image, and no-reference perceptual scores select among the render, restored observation, and high-frequency fused candidate. ClearGS then applies Full-Trajectory Repair Consolidation to revisit accepted repairs and preserve details introduced early. On GS2E and GSOTM, ClearGS achieves state-of-the-art overall performance, with consistent CLIP-IQA and MUSIQ gains and LPIPS reductions in most degradation settings, without paired sharp supervision or matched clean references.
Tags
Links
- Source: https://arxiv.org/abs/2609.31509v1
- Canonical: https://arxiv.org/abs/2609.31509v1
Trouble viewing inline? Open PDF directly →
Full Text
42,079 characters extracted from source content.
Expand or collapse full text
ClearGS: Reliability-Aware Gaussian Splatting from Handheld Videos Xuanzhi Liu Xinyi Wu Hang Pan Wensi Huang Zhenyao Wu Ruize Han Song Wang Abstract We present ClearGS for 3D Gaussian Splatting (3DGS) from handheld videos with uneven viewpoint coverage and mixed frame quality. Rather than selecting frames with binary decisions, ClearGS uses Reliability-aware View Allocation (RVA) to assign graded raw-supervision weights based on appearance reliability, degradation risk, and geometric utility, while weakly reactivating useful suppressed frames to maintain trajectory coverage. Since weighting cannot restore details lost to blur or distortion, ClearGS further introduces Render-Guided In-Video Restoration (RIVR). The current 3DGS render provides a pose-aligned structural candidate, a frozen no-reference restoration expert restores the corresponding raw video observation without any clean reference image, and no-reference perceptual scores select among the render, restored observation, and high-frequency fused candidate. ClearGS then applies Full-Trajectory Repair Consolidation to revisit accepted repairs and preserve details introduced early. On GS2E and GSOTM, ClearGS achieves state-of-the-art overall performance, with consistent CLIP-IQA and MUSIQ gains and LPIPS reductions in most degradation settings, without paired sharp supervision or matched clean references. 1Shenzhen University of Advanced Technology 2HONOR 3University of New South Wales 4Southern University of Science and Technology Figure 1: Traditional pipelines either train or select handheld-video frames without graded reliability control, then repair degraded views using nearby references or cross-view cues that may share the same artifacts. ClearGS instead uses RVA to separate appearance reliability from geometric utility, and applies no-reference RIVR to add selected repair evidence without matched clean references. Introduction Recent advances in 3D Gaussian Splatting (3DGS) enable efficient scene reconstruction and high-quality novel-view rendering (Kerbl et al. 2023), but they assume input views with reasonably consistent appearance and geometry. Handheld videos often break this assumption: frame quality changes along the trajectory, and viewpoint coverage is uneven. Training all frames with equal strength transfers blur, distortion, and inconsistent appearance into the scene. Existing frame- and view-selection methods retain or acquire views according to quality, coverage, uncertainty, or information gain (Polyzos et al. 2025; Xue et al. 2026; Chen et al. 2026b), but binary keep-or-drop decisions are too coarse for handheld video. A degraded frame may still provide useful geometric evidence; dropping it weakens coverage, while using it at full weight contaminates appearance. We therefore formulate handheld-video 3DGS as a mixed-quality supervision problem that separates appearance reliability from geometric utility. ClearGS first addresses this problem with Reliability-aware View Allocation (RVA). For each frame, RVA estimates appearance reliability, degradation risk, and geometric utility, then assigns an ordered raw-supervision weight. Reliable low-risk frames receive strong supervision, geometrically useful observations receive attenuated supervision, and unreliable low-utility frames are suppressed. To avoid coverage holes, RVA also weakly reactivates useful suppressed frames in under-covered trajectory regions, preserving viewpoint coverage without letting degraded appearance dominate the initial reconstruction. Weighted supervision alone cannot recover details already lost to blur or distortion. Existing restoration–distillation methods often depend on nearby RGB references, aggregated sparse-view features, or depth-guided correspondences (Wu et al. 2025a; Liu et al. 2024; Wu et al. 2025b; Yin et al. 2025; Cao et al. 2026). These cues are unreliable in handheld videos: neighboring frames may suffer from the same degradation, and corrupted appearance can be propagated across views. Render-Guided In-Video Restoration (RIVR) extends ClearGS from raw supervision to repaired evidence. For each target pose, RIVR renders the current scene as a pose-aligned structural candidate and restores the corresponding raw video observation with a frozen no-reference restoration expert, without using any clean reference image or paired sharp target. RIVR also constructs a detail-fused candidate by combining the rendered base with restored high-frequency details. It then selects among the render, restored observation, and fused candidate using no-reference perceptual scores, and appends the selected result as a pseudo observation for later distillation. Progressive distillation can weaken details introduced by repairs accepted early in the trajectory. ClearGS therefore performs Full-Trajectory Repair Consolidation, which reuses the latest accepted repair at each target pose without sampling new poses or generating new candidates. Together, RVA, RIVR, and Full-Trajectory Repair Consolidation preserve reliable captured evidence, recover video-grounded details, and maintain weakly observed regions. Experiments on GS2E (Li et al. 2025) and GSOTM (Seiskari et al. 2024) show state-of-the-art overall performance, with sharper boundaries and more stable textures under mixed-quality observations. Our contributions are threefold: • We formulate 3DGS reconstruction from handheld videos as a mixed-quality supervision problem that separates appearance reliability, degradation risk, and geometric utility. • We introduce RVA, which assigns graded raw-supervision weights and preserves trajectory coverage through weak reactivation without allowing degraded observations to dominate scene appearance. • We introduce RIVR, a no-reference in-video restoration and distillation strategy that selects repair evidence without clean references and consolidates accepted repairs through Full-Trajectory Repair Consolidation. Figure 2: Overview of ClearGS. RVA assigns frame-wise raw-supervision weights from appearance reliability, degradation risk, and geometric utility, and weakly reactivates useful suppressed frames in uncovered trajectory regions. RIVR selects among the current render, restored raw observation, and detail-fused candidate using no-reference quality scores, then distills accepted pseudo observations with real observations. Related Work Scene Reconstruction from Blurry Images and Handheld Videos Blur-aware radiance-field methods recover sharp scenes by learning spatially varying kernels, imposing physical priors, or estimating exposure-time camera trajectories (Ma et al. 2022; Lee et al. 2023a; Wang et al. 2023; Lee et al. 2023b). Recent 3DGS methods model blur through Gaussian covariance manipulation, per-pixel kernels, or continuous camera motion (Lee et al. 2024; Zhao et al. 2024; Peng et al. 2024; Lee et al. 2025), and handheld-video methods further compensate for rolling shutter and pose errors (Seiskari et al. 2024; Zhang et al. 2026). These methods explain degradation through image-formation or camera-motion models. ClearGS addresses a different failure mode: frame quality and viewpoint coverage vary along a fixed handheld trajectory. It controls the appearance influence of each frame while retaining observations that still provide useful geometric evidence. Frame and View Selection for Reconstruction Frame and active view selection methods retain informative observations or acquire new viewpoints according to image quality, coverage, uncertainty, or information gain (Polyzos et al. 2025; Xue et al. 2026; Chen et al. 2026b). Their goal is often to improve acquisition efficiency or expose under-observed geometry. In fixed handheld videos, however, binary selection cannot use the geometric value of a degraded frame without also accepting its unreliable appearance. RVA separates appearance reliability, degradation risk, and geometric utility, assigning graded raw-supervision weights rather than hard decisions. It also weakly reactivates useful suppressed frames to avoid trajectory coverage holes. 3DGS Recovery from Imperfect Observations Methods for sparse or imperfect observations stabilize 3DGS using depth priors, multi-view constraints, structural regularization, and Gaussian control (Chung et al. 2024; Xiong et al. 2023; Zhu et al. 2024; Li et al. 2024; Xu et al. 2024; Zhang et al. 2024; Park et al. 2025; Chen et al. 2026a; Song et al. 2025). Generative methods repair rendered views or add pseudo-observations (Wu et al. 2024; Liu et al. 2024; Wu et al. 2025b; Shen et al. 2026; Jiang et al. 2026; Li et al. 2026; Ni et al. 2025; Zhu et al. 2026), while recent 3DGS repair methods often condition restoration on nearby RGB references, sparse-view features, or depth-guided correspondences (Wu et al. 2025a; Yin et al. 2025; Cao et al. 2026). These cues can be unreliable in mixed-quality handheld videos, where neighboring frames may share the same degradation and corrupted appearance can propagate across views. RIVR instead restores the corresponding raw video observation with a frozen no-reference restoration expert, using no clean reference image or paired sharp target. The current scene render provides a pose-aligned structural candidate for selection and fusion, and accepted repairs are later reused through Full-Trajectory Repair Consolidation. Method Problem Definition We consider a handheld posed video sequence =(t,t)t=1T,V=\(I_t, π_t)\_t=1^T, (1) where tI_t is an input video frame and t π_t denotes its camera parameters. Such videos contain observations of heterogeneous quality: clear frames coexist with frames affected by motion blur, defocus, distortion, or other degradations. Moreover, viewpoint coverage is often uneven along the camera trajectory. A degraded frame may therefore be unsafe as appearance supervision while still providing the only observation of a weakly covered region. Our goal is to reconstruct a 3DGS scene ⋆G from V that faithfully renders both observed and novel views. During reconstruction, only the input video is available; paired sharp targets and high-quality reference images matched to the target pose are unavailable. The central challenge is to preserve the geometric coverage offered by degraded observations without directly writing their unreliable appearance into the scene representation. Table 1: RVA notation, parameter settings, and supervision policy. The standard supervision conditions are evaluated from top to bottom. Afterward, trajectory coverage compensation may reactivate a suppressed frame in an uncovered trajectory bin with a fixed weak-supervision weight. Part Quantity / role Definition or condition Parameters / weight Appearance ir_i ordered quality cues: sharpness, structural clarity, texture reliability, artifact cleanliness, exposure quality R=(0.30,0.25,0.20,0.15,0.10)w_R=(0.30,0.25,0.20,0.15,0.10) ib_i ordered degradation cues: motion blur, defocus, ghosting, exposure failure, unreliable texture, artifacts B=(0.25,0.15,0.20,0.15,0.15,0.10)w_B=(0.25,0.15,0.20,0.15,0.15,0.10) RiR_i appearance reliability score Ri=R⊤iR_i=w_R r_i BiB_i degradation risk score Bi=B⊤iB_i=w_B b_i A appearance-reliable frame set τR=0.70,τB=0.60 _R=0.70,\ _B=0.60 Geometry dijcd_ij^c normalized camera-center distance between frames i and j Euclidean distance normalized to [0,1][0,1] dijvd_ij^v normalized viewing-direction difference between frames i and j viewing-direction angle normalized to [0,1][0,1] DiD_i Di=maxj∈(γcdijc+γvdijv) D_i= _j ( _cd_ij^c+ _vd_ij^v ) (γc,γv)=(0.65,0.35)( _c, _v)=(0.65,0.35) OiO_i Oi=maxj∈|i∩j||i∪j| O_i= _j |P_i _j | |P_i _j | – GiG_i geometric utility score (α,β,η)=(0.45,0.35,0.30)(α,β,η)=(0.45,0.35,0.30) Trajectory compensation Selection coefficients coefficients in Eq. (6) (λG,λR,λB,λZ)=(0.35,0.25,0.30,0.10)( _G, _R, _B, _Z)=(0.35,0.25,0.30,0.10) Reactivated supervision suppressed frame selected as im∗i_m in an uncovered trajectory bin wim∗=0.12w_i_m =0.12 Supervision assignment Strong Ri≥0.70,Bi≤0.60R_i≥ 0.70,\ B_i≤ 0.60 wi=0.80+0.20Riw_i=0.80+0.20R_i Geometrically attenuated Gi≥0.60,Bi≤0.80G_i≥ 0.60,\ B_i≤ 0.80 wi=0.30+0.30(0.35Ri+0.65Gi)w_i=0.30+0.30(0.35R_i+0.65G_i) Auxiliary attenuated Ri≥0.35,Gi≥0.25,Bi≤0.75R_i≥ 0.35,\ G_i≥ 0.25,\ B_i≤ 0.75 wi=0.30+0.30(0.50Ri+0.50Gi)w_i=0.30+0.30(0.50R_i+0.50G_i) Coverage-reactivated initially suppressed but selected as im∗i_m wi=0.12w_i=0.12 Suppressed otherwise and not selected for coverage compensation wi=0w_i=0 Overview ClearGS follows a reliability-first principle for handheld video reconstruction: we first establish a reliable initial scene before introducing repaired observations. RVA evaluates frames from appearance and geometric perspectives, allocates reliability-aware supervision, and preserves necessary trajectory coverage to initialize 0G_0. Starting from 0G_0, RIVR forms three target-pose candidates: the current 3DGS render, the restored raw video frame, and a fused result that combines the rendered base with restored high-frequency details. The selected candidate is appended as an additional pseudo observation and optimized together with the original video supervision, complementing rather than replacing the captured frames. Finally, full-trajectory repair consolidation gathers the accepted pseudo observations across the trajectory and performs a final joint distillation pass, producing the optimized scene ⋆G . Reliability-aware View Allocation (RVA) Handheld video reconstruction is influenced by both the appearance reliability of individual frames and their geometric utility for scene reconstruction. Frames with reliable appearance provide stronger photometric supervision, while frames captured from complementary viewpoints provide additional geometric constraints. Therefore, ClearGS evaluates each frame from both appearance and geometric perspectives and assigns reliability-aware supervision weights, instead of simply retaining or discarding frames based on a single criterion (Polyzos et al. 2025; Xue et al. 2026; Chen et al. 2026b). Appearance Evidence We first evaluate whether a frame provides reliable appearance information. For each input frame iI_i, RVA uses a frozen Qwen3-VL (Bai et al. 2025) as an image-based evaluator. Qwen3-VL predicts appearance quality cues and degradation patterns, which are aggregated into two complementary scores. Quality-related cues, including sharpness, structural clarity, texture reliability, artifact cleanliness, and exposure quality, are combined into a reliability score RiR_i, indicating whether the frame can provide trustworthy appearance supervision. Degradation patterns, including motion blur, defocus, ghosting, exposure failure, unreliable texture, and artifacts, are combined into a degradation risk score BiB_i, indicating the likelihood of introducing unreliable appearance into the reconstruction. Frames with high RiR_i and low BiB_i are considered appearance-reliable frames A: =j∣Rj≥τR,Bj≤τB,A=\j R_j≥ _R,\ B_j≤ _B\, (2) where τR _R and τB _B are predefined thresholds for reliability and degradation risk, respectively. The score aggregation weights and threshold values are summarized in Table 1. Geometric Evidence Besides appearance information, ClearGS considers the geometric contribution of each frame to scene reconstruction. A frame with degraded appearance may still provide valuable observations from complementary viewpoints. We estimate geometric utility based on the COLMAP sparse reconstruction and camera poses. Given the reliable frame set A obtained from appearance assessment, we construct a reference coverage by merging their visible 3D points: rel=⋃j∈j,P_rel= _j P_j, (3) where jP_j denotes the set of COLMAP 3D points observed by frame j. This coverage represents scene regions supported by reliable appearance observations. For each frame i, we measure geometric utility using three complementary cues: coverage gain CiC_i, pose diversity DiD_i, and visibility redundancy OiO_i. Coverage gain measures additional 3D points observed by frame i beyond relP_rel: Ci=∑p∈i∖rel1np∑p∈i1np,C_i= _p _i _rel 1 n_p _p _i 1 n_p, (4) where npn_p denotes the number of COLMAP images observing point p, assigning larger weights to sparsely observed points. Pose diversity measures the complementarity of the camera pose of frame i with reliable observations, while visibility redundancy measures its overlap with existing coverage. The geometric utility is computed as: Gi=clip[0,1](αCi+βDi−ηOi),G_i=clip_[0,1] (α C_i+β D_i-η O_i ), (5) where α, β, and η balance the three geometric cues. A higher GiG_i indicates that the frame provides complementary geometric evidence and receives higher priority during supervision allocation. The formulations of DiD_i and OiO_i, together with their corresponding weights, are summarized in Table 1. Trajectory Coverage Compensation However, independent frame-wise decisions may suppress neighboring frames in the same trajectory region, causing local coverage gaps. To preserve sufficient camera coverage during scene initialization, RVA further performs trajectory coverage compensation. Specifically, the trajectory is partitioned into pose-aware bins according to camera motion. If a bin already contains active frames, it remains unchanged. Otherwise, RVA reactivates one suppressed frame im∗i_m : im∗=argmaxi∈m(λGGi+λRRi+λB(1−Bi)+λZZi),i_m = _i _m ( _GG_i+ _RR_i+ _B(1-B_i)+ _ZZ_i ), (6) where mS_m denotes the suppressed frames in the uncovered trajectory bin. ZiZ_i denotes the trajectory centrality score, where larger values favor frames closer to the bin center. The trajectory-compensation coefficients and weak reactivation weight are summarized in Table 1. Finally, RVA assigns each frame a supervision role according to Table 1. The resulting supervision set, including both allocated frames and trajectory-compensated frames, is used to initialize 0G_0. Render-Guided In-Video Restoration (RIVR) After RVA builds the initial scene 0G_0, some details from degraded frames may still be missing or weak in the reconstruction. RIVR addresses this problem with no-reference in-video restoration. It does not require a matched clean image, a paired sharp target, nearby RGB references, aggregated cross-view features, or explicit geometric correspondences (Wu et al. 2025a; Liu et al. 2024; Wu et al. 2025b; Yin et al. 2025; Cao et al. 2026). Instead, it uses two signals available at the same target pose: the current 3DGS render and the corresponding raw video observation. For each target frame t, RIVR renders the current scene as tR_t and retrieves the raw observation tO_t. The raw observation, rather than the rendered image, is processed by a frozen Restormer motion-deblurring expert (Zamir et al. 2022), producing tD_t. The render provides pose-aligned scene structure, but it is not treated as a clean reference. The restored observation provides target-view appearance evidence recovered from the original video frame. RIVR then forms the candidate set summarized in Table 2. Table 2: RIVR candidates for no-reference in-video restoration. Candidate Construction Role Render tR_t Pose-aligned structural candidate Restored tD_t Restoration of the raw target observation Fused t=t+t−ℬσ(t)F_t=R_t+D_t-B_σ(D_t) Render base with restored high-frequency detail Because clean references are unavailable during reconstruction, RIVR selects among the candidates with a reference-free rule. MUSIQ (Ke et al. 2021) is used as the primary score because it responds to blur, detail loss, and unnatural local structure, while CLIP-IQA is used only when MUSIQ scores are close. The selected candidate is appended as a pseudo observation for later distillation. Thus, restoration evidence is used only after being chosen from render, restored, and fused candidates, without assuming any clean target image or cross-view reference. Joint Distillation Following restoration-guided reconstruction methods (Wu et al. 2025a; Yin et al. 2025; Cao et al. 2026), ClearGS distills accepted RIVR outputs with the active real views selected by RVA. At round r, the training set is (r)=real∪⋃ℓ=1rpseudo(ℓ).D^(r)=D_real∪ _ =1^rD_pseudo^( ). (7) Repaired views are appended rather than used to replace real observations, keeping captured video evidence while adding selected restoration details. Since later rounds may weaken repairs accepted early, ClearGS applies Full-Trajectory Repair Consolidation after all frames are processed. It reuses the accepted repairs from the whole trajectory, keeps only the latest repair for repeated target frames, and introduces no new poses or repair candidates. Experiment Experimental Setup We evaluate ClearGS on GS2E (Li et al. 2025) and Gaussian Splatting on the Move (GSOTM) (Seiskari et al. 2024). GS2E contains ten scenes at each of three synthetic blur levels: slight, medium, and severe. GSOTM complements this protocol with four camera-motion degradation settings: motion blur (MB), rolling shutter (RS), their combination (MB+RS), and pose noise (PN). Every method receives the same degraded posed observations. Neither sharp ground-truth images nor matched high-quality reference images are available during view allocation, reconstruction, repair selection, or distillation; sharp targets are used only for final reference-based evaluation. Table 3: Perceptual comparison on GS2E and GSOTM at observed input poses. GS2E averages ten scenes per blur level, and GSOTM averages four scenes per degradation type. LPIPS uses sharp targets; CLIP-IQA and MUSIQ are no-reference metrics. Best and second-best results are highlighted. Method LPIPS↓ CLIP-IQA↑ MUSIQ↑ GS2E GSOTM GS2E GSOTM GS2E GSOTM Slight Medium Severe MB RS MB+RS PN Slight Medium Severe MB RS MB+RS PN Slight Medium Severe MB RS MB+RS PN All-view G0G_0 0.397 0.414 0.430 0.323 0.334 0.419 0.476 0.266 0.261 0.256 0.194 0.212 0.196 0.227 34.82 32.02 30.11 36.75 50.25 34.51 34.06 Select-G G0G_0 (Ours) 0.386 0.401 0.415 0.285 0.333 0.393 0.472 0.266 0.256 0.261 0.201 0.211 0.187 0.224 36.53 34.51 34.22 42.31 50.55 39.46 35.42 DiFiX3D (Wu et al. 2025a) 0.349 0.372 0.388 0.224 0.224 0.322 0.249 0.258 0.255 0.253 0.231 0.253 0.213 0.230 41.53 39.03 36.93 41.64 59.77 39.94 51.64 GeoQuery (Cao et al. 2026) 0.350 0.373 0.389 0.231 0.224 0.332 0.249 0.256 0.253 0.252 0.230 0.250 0.216 0.231 41.33 38.70 37.20 41.86 59.50 40.02 51.27 SyncFix (Li et al. 2026) 0.353 0.377 0.395 0.253 0.227 0.334 0.250 0.253 0.246 0.242 0.225 0.249 0.208 0.227 39.99 36.76 34.44 37.90 59.00 35.67 50.26 ClearGS (Ours) 0.333 0.342 0.350 0.193 0.225 0.302 0.197 0.288 0.287 0.293 0.314 0.296 0.243 0.267 44.67 43.57 43.18 62.03 64.25 52.64 59.07 Table 4: Ablation of RVA coverage allocation. S/M/V are GS2E blur levels; MB/RS/MB+RS/PN are GSOTM degradation settings. Variant PSNR↑ SSIM↑ LPIPS↓ GS2E GSOTM GS2E GSOTM GS2E GSOTM S M V MB RS MB+RS PN S M V MB RS MB+RS PN S M V MB RS MB+RS PN RVA w/o cov. 20.123 19.538 19.499 21.621 18.010 17.624 18.413 0.607 0.592 0.589 0.730 0.530 0.534 0.573 0.388 0.413 0.415 0.290 0.339 0.387 0.472 Full RVA 19.934 20.076 20.011 22.545 18.124 18.010 18.458 0.606 0.609 0.608 0.739 0.530 0.541 0.573 0.386 0.401 0.415 0.285 0.333 0.393 0.472 Effect of Reliability-aware View Allocation We isolate the reconstruction stage by comparing all-view 0G_0 with RVA-selected 0G_0. As shown in Figure 3, all-view supervision propagates blur into thin structures and object boundaries, while RVA yields sharper contours and more stable local appearance. The initial-scene rows in Table 3 show the same trend: RVA-selected 0G_0 reduces LPIPS by 0.011, 0.013, and 0.015 and improves MUSIQ by 1.71, 2.49, and 4.11 from slight to severe blur. CLIP-IQA remains comparable, with an equal score under slight blur and a small gain under severe blur. These results show that RVA reduces blur-related degradation and provides a cleaner initialization for repair. Figure 3: Qualitative comparison of renderings at observed input camera poses produced by all-view 3DGS and the Select-view initialization of ClearGS. Red boxes highlight regions where our reconstruction better preserves thin structures, object boundaries, and local appearance relative to all-view supervision. Quantitative Comparison Table 3 reports perceptual results on GS2E and GSOTM, together with all-view and selected-view initial scenes to separate view allocation from repair. For a fair comparison, all methods use the same input sequences and are evaluated at the same observed input poses against the same sharp targets. Under this protocol, ClearGS improves over DiFiX3D on all GS2E blur levels, reducing LPIPS by 0.016, 0.030, and 0.038 and increasing CLIP-IQA by 0.030, 0.032, and 0.040 from slight to severe blur. MUSIQ gains also grow with blur strength, from 3.14 to 6.25. On GSOTM, ClearGS achieves the best CLIP-IQA and MUSIQ in all four settings, and its LPIPS is best in three settings and within 0.001 of the best score under rolling shutter. Qualitative Comparison Figure 4: Qualitative comparison of repair outputs on two GS2E scenes. Each row shows the same camera pose. From left to right, the panels show the common pre-repair 3DGS rendering, the repairs produced by DiFiX3D and GeoQuery, and the repair selected by ClearGS. Table 5: Ablation of RIVR candidate selection. “w/o Obs.” replaces the raw video observation OtO_t with the current render RtR_t as the input to Restormer, while “w/o FtF_t” removes the high-frequency fused candidate. The remaining repair-distillation pipeline is unchanged. Metric Variant GS2E GSOTM S M V MB RS MB+RS PN LPIPS↓ w/o Obs. 0.385 0.401 0.414 0.247 0.294 0.358 0.398 w/o FtF_t 0.336 0.348 0.365 0.171 0.213 0.298 0.196 Full RIVR 0.333 0.342 0.350 0.193 0.225 0.302 0.197 CLIP-IQA↑ w/o Obs. 0.283 0.270 0.279 0.217 0.239 0.209 0.258 w/o FtF_t 0.280 0.275 0.282 0.263 0.280 0.221 0.269 Full RIVR 0.288 0.287 0.293 0.314 0.296 0.243 0.267 MUSIQ↑ w/o Obs. 37.15 35.59 34.98 47.37 57.30 46.20 47.13 w/o FtF_t 43.43 42.12 40.67 51.64 63.30 45.53 58.98 Full RIVR 44.67 43.57 43.18 62.03 64.25 52.64 59.07 Figure 5: Qualitative comparison on five GS2E scenes. From left to right: the input 3DGS, DiFiX3D, GeoQuery, SyncFix, ClearGS, and the ground truth. ClearGS recovers sharper object boundaries and local details while preserving the scene appearance. Figure 4 compares repair outputs at the same observed camera poses, while Figure 5 compares the final renderings with the input 3DGS, existing restoration baselines, and sharp ground truth. The pre-repair renderings retain visible blur and weakened local structures, and the restoration baselines improve sharpness but can still leave residual artifacts or less stable scene details. In contrast, ClearGS produces cleaner object boundaries and more coherent local textures in the shown examples. These visual differences align with the quantitative results in Table 3, where ClearGS obtains the best overall perceptual scores and larger gains under stronger degradations. Ablation Study Coverage Allocation in RVA Table 4 evaluates the coverage top-up step in RVA. On GS2E, full RVA improves PSNR by 0.538 and 0.512 and SSIM by 0.017 and 0.019 under medium and severe blur, where reliability weighting is more likely to suppress useful but degraded trajectory regions. It also lowers LPIPS under slight and medium blur, while matching the no-top-up variant under severe blur. Under slight blur, the small PSNR drop and nearly unchanged SSIM suggest limited gain when most views remain reliable. On GSOTM, the effect is mixed but improves several settings, including RS and PN. These results show that coverage top-up complements reliability weighting by recovering useful trajectory evidence under stronger degradation. RIVR Candidate Selection Table 5 ablates the RIVR candidate pool. The full model selects among the current render RtR_t, the no-reference Restormer output from the aligned video observation OtO_t, and the high-frequency fusion candidate FtF_t. In w/o OtO_t, Restormer is still used, but its input is changed from OtO_t to RtR_t; thus the ablation removes video evidence rather than the repair module. Removing this candidate consistently hurts LPIPS, CLIP-IQA, and MUSIQ on both datasets, with large MUSIQ drops on GSOTM motion blur and pose noise (14.66 and 11.94). Removing FtF_t generally lowers MUSIQ and CLIP-IQA, indicating that the fused candidate contributes perceptually preferred high-frequency details. Conclusion and Discussion We presented ClearGS for 3DGS reconstruction from handheld videos with mixed frame quality and uneven trajectory coverage. ClearGS assigns graded supervision through RVA, separating visual reliability, degradation risk, and geometric utility while preserving coverage in weakly observed trajectory regions. It then performs no-reference in-video restoration: the current render serves as a pose-aligned structural anchor, and the corresponding raw video frame provides the evidence to be restored, without requiring matched clean references. A final consolidation pass revisits accepted repairs across the trajectory to preserve details learned early. Experiments on GS2E and GSOTM show improved perceptual quality, sharper boundaries, and more stable textures. Future work will study degradation-specific restoration and geometric checks for repaired evidence. References Bai et al. (2025) S. Bai, Y. Cai, R. Chen, K. Chen, X. Chen, Z. Cheng, L. Deng, W. Ding, C. Gao, C. Ge, et al. Qwen3-vl technical report. arXiv preprint arXiv:2511.21631. Cited by: Appearance Evidence. Cao et al. (2026) X. Cao, Y. Li, Y. Zhang, J. Song, C. Yan, W. Li, and L. Duan GeoQuery: geometry-query diffusion for sparse-view reconstruction. arXiv preprint arXiv:2605.12399. Cited by: Introduction, 3DGS Recovery from Imperfect Observations, Render-Guided In-Video Restoration (RIVR), Joint Distillation, Table 3. Chen et al. (2026a) K. Chen, Y. Zhong, Z. Li, J. Lin, Y. Chen, M. Qin, and H. Wang Quantifying and alleviating co-adaptation in sparse-view 3d gaussian splatting. Advances in Neural Information Processing Systems 38, p. 115939–115968. Cited by: 3DGS Recovery from Imperfect Observations. Chen et al. (2026b) T. Chen, A. Dai, M. Adang, G. Gao, and M. Schwager Coverage optimization for camera view selection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 19443–19451. Cited by: Introduction, Frame and View Selection for Reconstruction, Reliability-aware View Allocation (RVA). Chung et al. (2024) J. Chung, J. Oh, and K. M. Lee Depth-regularized optimization for 3d gaussian splatting in few-shot images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 811–820. Cited by: 3DGS Recovery from Imperfect Observations. Jiang et al. (2026) C. Jiang, Y. Chen, Q. Zhang, J. Song, S. Xu, D. Yeung, and J. Deng FreeScale: scaling 3d scenes via certainty-aware free-view generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 330–340. Cited by: 3DGS Recovery from Imperfect Observations. Ke et al. (2021) J. Ke, Q. Wang, Y. Wang, P. Milanfar, and F. Yang Musiq: multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, p. 5148–5157. Cited by: Render-Guided In-Video Restoration (RIVR). Kerbl et al. (2023) B. Kerbl, G. Kopanas, T. Leimkühler, G. Drettakis, et al. 3d gaussian splatting for real-time radiance field rendering.. ACM Trans. Graph. 42 (4), p. 139–1. Cited by: Introduction. Lee et al. (2024) B. Lee, H. Lee, X. Sun, U. Ali, and E. Park Deblurring 3d gaussian splatting. In European Conference on Computer Vision, p. 127–143. Cited by: Scene Reconstruction from Blurry Images and Handheld Videos. Lee et al. (2023a) D. Lee, M. Lee, C. Shin, and S. Lee Dp-nerf: deblurred neural radiance field with physical scene priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 12386–12396. Cited by: Scene Reconstruction from Blurry Images and Handheld Videos. Lee et al. (2023b) D. Lee, J. Oh, J. Rim, S. Cho, and K. M. Lee Exblurf: efficient radiance fields for extreme motion blurred images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 17639–17648. Cited by: Scene Reconstruction from Blurry Images and Handheld Videos. Lee et al. (2025) J. Lee, D. Kim, D. Lee, S. Cho, M. Lee, W. Lee, T. Kim, D. Wee, and S. Lee Comogaussian: continuous motion-aware gaussian splatting from motion-blurred images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 26415–26424. Cited by: Scene Reconstruction from Blurry Images and Handheld Videos. Li et al. (2026) D. Li, A. Yadav, C. Peng, R. Chellappa, and A. Bhattad SyncFix: fixing 3d reconstructions via multi-view synchronization. arXiv preprint arXiv:2604.11797. Cited by: 3DGS Recovery from Imperfect Observations, Table 3. Li et al. (2024) J. Li, J. Zhang, X. Bai, J. Zheng, X. Ning, J. Zhou, and L. Gu Dngaussian: optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 20775–20785. Cited by: 3DGS Recovery from Imperfect Observations. Li et al. (2025) Y. Li, C. Feng, Z. Tang, K. Deng, W. Yu, Y. Tian, and L. Yuan GS2E: gaussian splatting is an effective data generator for event stream generation. In Advances in Neural Information Processing Systems, Vol. 38. Cited by: Introduction, Experimental Setup. Liu et al. (2024) X. Liu, C. Zhou, and S. Huang 3dgs-enhancer: enhancing unbounded 3d gaussian splatting with view-consistent 2d diffusion priors. Advances in Neural Information Processing Systems 37, p. 133305–133327. Cited by: Introduction, 3DGS Recovery from Imperfect Observations, Render-Guided In-Video Restoration (RIVR). Ma et al. (2022) L. Ma, X. Li, J. Liao, Q. Zhang, X. Wang, J. Wang, and P. V. Sander Deblur-nerf: neural radiance fields from blurry images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 12861–12870. Cited by: Scene Reconstruction from Blurry Images and Handheld Videos. Ni et al. (2025) J. Ni, Y. Chen, Z. Yang, Y. Liu, R. Lu, S. Zhu, and S. Huang G4Splat: geometry-guided gaussian splatting with generative prior. arXiv preprint arXiv:2510.12099. Cited by: 3DGS Recovery from Imperfect Observations. Park et al. (2025) H. Park, G. Ryu, and W. Kim Dropgaussian: structural regularization for sparse-view gaussian splatting. In Proceedings of the computer vision and pattern recognition conference, p. 21600–21609. Cited by: 3DGS Recovery from Imperfect Observations. Peng et al. (2024) C. Peng, Y. Tang, Y. Zhou, N. Wang, X. Liu, D. Li, and R. Chellappa Bags: blur agnostic gaussian splatting through multi-scale kernel modeling. In European Conference on Computer Vision, p. 293–310. Cited by: Scene Reconstruction from Blurry Images and Handheld Videos. Polyzos et al. (2025) K. D. Polyzos, A. Bacharis, S. Madhuvarasu, N. Papanikolopoulos, and T. Javidi Activeinitsplat: how active image selection helps gaussian splatting. arXiv preprint arXiv:2503.06859. Cited by: Introduction, Frame and View Selection for Reconstruction, Reliability-aware View Allocation (RVA). Seiskari et al. (2024) O. Seiskari, J. Ylilammi, V. Kaatrasalo, P. Rantalankila, M. Turkulainen, J. Kannala, E. Rahtu, and A. Solin Gaussian splatting on the move: blur and rolling shutter compensation for natural camera motion. In European conference on computer vision, p. 160–177. Cited by: Introduction, Scene Reconstruction from Blurry Images and Handheld Videos, Experimental Setup. Shen et al. (2026) Y. Shen, Y. Yang, K. Zhou, and T. Shao Sparse-to-complete: from sparse image captures to complete 3d scenes. arXiv preprint arXiv:2605.05664. Cited by: 3DGS Recovery from Imperfect Observations. Song et al. (2025) M. Song, X. Lin, D. Zhang, H. Li, X. Li, B. Du, and L. Qi D2^2GS: depth-and-density guided gaussian splatting for stable and accurate sparse-view reconstruction. arXiv preprint arXiv:2510.08566. Cited by: 3DGS Recovery from Imperfect Observations. Wang et al. (2023) P. Wang, L. Zhao, R. Ma, and P. Liu Bad-nerf: bundle adjusted deblur neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 4170–4179. Cited by: Scene Reconstruction from Blurry Images and Handheld Videos. Wu et al. (2025a) J. Z. Wu, Y. Zhang, H. Turki, X. Ren, J. Gao, M. Z. Shou, S. Fidler, Z. Gojcic, and H. Ling Difix3d+: improving 3d reconstructions with single-step diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 26024–26035. Cited by: Introduction, 3DGS Recovery from Imperfect Observations, Render-Guided In-Video Restoration (RIVR), Joint Distillation, Table 3. Wu et al. (2024) R. Wu, B. Mildenhall, P. Henzler, K. Park, R. Gao, D. Watson, P. P. Srinivasan, D. Verbin, J. T. Barron, B. Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 21551–21561. Cited by: 3DGS Recovery from Imperfect Observations. Wu et al. (2025b) S. Wu, C. Xu, B. Huang, A. Geiger, and A. Chen Genfusion: closing the loop between reconstruction and generation via videos. In Proceedings of the Computer Vision and Pattern Recognition Conference, p. 6078–6088. Cited by: Introduction, 3DGS Recovery from Imperfect Observations, Render-Guided In-Video Restoration (RIVR). Xiong et al. (2023) H. Xiong, S. Muttukuru, R. Upadhyay, P. Chari, and A. Kadambi Sparsegs: real-time 360 sparse view synthesis using gaussian splatting. arXiv e-prints, p. arXiv–2312. Cited by: 3DGS Recovery from Imperfect Observations. Xu et al. (2024) W. Xu, H. Gao, S. Shen, R. Peng, J. Jiao, and R. Wang Mvpgs: excavating multi-view priors for gaussian splatting from sparse input views. In European Conference on Computer Vision, p. 203–220. Cited by: 3DGS Recovery from Imperfect Observations. Xue et al. (2026) S. Xue, J. Dill, D. Ahuja, F. Dellaert, P. Tsiotras, and D. Xu Uncertainty-driven 3d gaussian splatting active mapping via anisotropic visibility field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 5014–5026. Cited by: Introduction, Frame and View Selection for Reconstruction, Reliability-aware View Allocation (RVA). Yin et al. (2025) X. Yin, Q. Zhang, J. Chang, Y. Feng, Q. Fan, X. Yang, C. Pun, H. Zhang, and X. Cun Gsfixer: improving 3d gaussian splatting with reference-guided video diffusion priors. arXiv preprint arXiv:2508.09667. Cited by: Introduction, 3DGS Recovery from Imperfect Observations, Render-Guided In-Video Restoration (RIVR), Joint Distillation. Zamir et al. (2022) S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, and M. Yang Restormer: efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 5728–5739. Cited by: Render-Guided In-Video Restoration (RIVR). Zhang et al. (2024) J. Zhang, J. Li, X. Yu, L. Huang, L. Gu, J. Zheng, and X. Bai Cor-gs: sparse-view 3d gaussian splatting via co-regularization. In European conference on computer vision, p. 335–352. Cited by: 3DGS Recovery from Imperfect Observations. Zhang et al. (2026) Q. Zhang, D. Rozumny, F. Girlanda, S. Karaoglu, M. Pollefeys, T. Gevers, and M. R. Oswald Unblur-slam: dense neural slam for blurry inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 352–362. Cited by: Scene Reconstruction from Blurry Images and Handheld Videos. Zhao et al. (2024) L. Zhao, P. Wang, and P. Liu Bad-gaussians: bundle adjusted deblur gaussian splatting. In European Conference on Computer Vision, p. 233–250. Cited by: Scene Reconstruction from Blurry Images and Handheld Videos. Zhu et al. (2026) L. Zhu, M. Narayana, M. Stary, W. Hutchcroft, G. Wetzstein, and I. Armeni GaussFusion: improving 3d reconstruction in the wild with a geometry-informed video generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 15432–15442. Cited by: 3DGS Recovery from Imperfect Observations. Zhu et al. (2024) Z. Zhu, Z. Fan, Y. Jiang, and Z. Wang Fsgs: real-time few-shot view synthesis using gaussian splatting. In European conference on computer vision, p. 145–163. Cited by: 3DGS Recovery from Imperfect Observations.