---

title: >- [Paper Note] PriSplat: Propagating Reliable Multi-view Information for Distractor-Free 3DGS description: >- [ECCV2026][3D Vision][3DGS] PriSplat reframes distractor-free 3DGS from naive zero-masking to 3D-aware multi-view harmonization using repurposed NVS priors and geometry-aware support view selection, eliminating floaters and achieving state-of-the-art reconstruction fidelity. tags: - ECCV2026 - 3D Vision - 3DGS - Distractor-free Reconstruction - Multi-view Inpainting - Novel View Synthesis date: 2026-09-19 content_hash: faccdb56db636730

PriSplat: Propagating Reliable Multi-view Information for Distractor-Free 3DGS

Conference: ECCV 2026
Paper: ECCV Paper
Code: https://github.com/yun-seo/PriSplat.git
Area: 3D Vision
Keywords: 3D Gaussian Splatting, distractor-free reconstruction, multi-view information propagation, 3D-aware inpainting, novel view synthesis prior

TL;DR

PriSplat moves beyond the naive zero-masking paradigm in distractor-free 3DGS by repurposing a large-scale novel view synthesis prior into a 3D-aware inpainting harmonizer guided by geometry-aware support view selection and mask-aware fast-weight updates, achieving a state-of-the-art mean PSNR of 23.11 dB on the On-the-go benchmark while effectively eliminating floaters.

Background & Motivation

3D Gaussian Splatting (3DGS) has rapidly become a standard representation for high-fidelity novel view synthesis (NVS), driven by its explicit primitive representation and fast rasterization pipeline. However, standard 3DGS relies fundamentally on the assumption that the observed scene is strictly static across all training views. In unconstrained in-the-wild photo collections, transient distractors—such as moving pedestrians, pets, and passing vehicles—frequently corrupt the image captures. These non-static entities violate multi-view photometric consistency, injecting severe gradient noise into the optimization and deteriorating the underlying geometric reconstruction.

To counter these transient disturbances, existing distractor-free 3DGS methods predominantly follow a "residual detection and loss masking" paradigm. In these approaches, transient regions are localized via semantic segmentation or photometric residual heuristics, and the loss over the identified distractor pixels is suppressed to zero during optimization. However, while zero-masking successfully blocks corrupted gradients, it simultaneously removes crucial multi-view supervisory constraints over the occluded static background. As a result, the unobserved regions suffer from severe under-supervision and per-view overfitting, manifesting as conspicuous floaters, ghosting artifacts, and blurred or collapsed geometries in novel synthesized viewpoints. While a few recent techniques attempt to fill these masked regions using 2D diffusion inpainting models, pure 2D priors lack multi-view spatial awareness and often hallucinate cross-view inconsistent structures, corrupting 3D optimization.

This paper tackles the challenge by recognizing that unconstrained multi-view collections possess abundant cross-view redundancy: a background region obscured by a transient distractor in one camera is often cleanly observable from other viewpoints. Furthermore, novel view synthesis itself is inherently a reference-based inpainting task conditioned on camera poses. Instead of relying on unconstrained 2D priors or leaving masked areas unsupervised, one can repurpose a multi-view NVS prior to propagate reliable background information from unoccluded views into occluded target regions. Core idea: repurpose a large-scale NVS prior into a 3D-aware multi-view inpainting harmonizer with mask-aware fast-weight updates to prevent distractor leakage, guided by Fisher-based information density and spatio-angular support view selection to provide dense, geometrically consistent pseudo-ground truth for 3DGS.

Method

Overall Architecture

PriSplat tightly integrates a 3D-aware multi-view harmonization mechanism into the 3DGS optimization pipeline, shifting from passive mask suppression to active multi-view evidence propagation. The pipeline operates across three interconnected stages: online mask prediction, geometry-aware support view selection, and multi-view harmonization.

During the initial training iterations (0 to 10k), a lightweight online MLP predicts transient distractor masks, and 3DGS optimizes purely on unmasked pixels to establish a stable static geometry scaffold without fitting dynamic elements. In the intermediate phase (10k to 20k iterations), the multi-view harmonization pipeline is activated: for each target view containing masked occlusions, the system searches for optimal support views that cleanly capture the occluded background using an information density and pose compatibility metric. The harmonizer then adapts its scene memory through mask-aware fast-weight updates and synthesizes geometrically consistent, full-resolution pseudo-ground truth images. 3DGS is subsequently supervised by dense reconstruction losses on these harmonized views to suppress floaters. Finally, in the refinement phase (20k to 30k iterations), the pipeline returns to standard masked supervision to polish high-frequency appearance from clean original observations.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["In-the-wild Multi-view Input Images"] --> B["Dynamic Distractor Mask Prediction<br/>Lightweight online MLP estimates transient regions"]
    B --> C["Stage 1: Mask Warm-up Optimization<br/>Coarse 3DGS geometry optimized on unmasked pixels"]
    C --> D["Stage 2: Geometry-aware Support View Selection<br/>Fisher information density + Spatial distance + Angular constraints"]
    D --> E["Stage 2: Multi-view Harmonizer Inpainting<br/>Mask-aware fast-weight update + Cross-view token aggregation"]
    E --> F["Dense Pseudo-GT Supervision for 3DGS<br/>Re-establishes dense constraints to eliminate floaters"]
    F --> G["Stage 3: Masked Refinement<br/>Converges to clean distractor-free 3D Gaussians"]

Key Designs

1. Geometry-aware Support View Selection: Joint Fisher Information Density and Spatio-Angular Filtering

Selecting support views based purely on camera proximity (e.g., Euclidean KNN) frequently fails because neighboring cameras often observe the exact same moving distractor, or feature extreme grazing angles that induce projection distortions. To supply the harmonizer with genuinely informative cross-view evidence, PriSplat designs a geometry-grounded selection metric.

For a target view \(t\), let \(\mathcal{G}_t^{\mathrm{rel}}\) denote the subset of 3D Gaussians whose splats project inside the masked occlusion region of \(t\). For any candidate reference view \(s \in \mathcal{S}_t\), its utility depends on how reliably and sensitively it supervises these occluded Gaussians. The method quantifies observation reliability via a Fisher information score computed from 3DGS gradients:

\[r_{t,s} = \frac{1}{|\mathcal{G}_t^{\mathrm{rel}}|} \sum_{k \in \mathcal{G}_t^{\mathrm{rel}}} \mathbb{I} \left[ g_k \text{ is well-observed in view } s \right]\]

where the indicator function \(\mathbb{I}[\cdot]\) evaluates the sensitivity of rendered pixels in view \(s\) to perturbations in Gaussian parameter \(\theta_k\) via squared gradient norms. To prevent severe parallax distortion and extreme viewing angles, the density score is modulated by spatial distance decay \(w_{\mathrm{dist}}(t, s) = \exp(-\|c_t - c_s\|_2 / \sigma_{\mathrm{dist}})\) and viewing angle alignment \(w_{\mathrm{ang}}(t, s) = \max(0, v_t \cdot v_s)^{k_{\mathrm{ang}}}\), yielding a combined reliability score \(\rho_{t,s} = r_{t,s} \cdot w_{\mathrm{dist}}(t,s) \cdot w_{\mathrm{ang}}(t,s)\). Support candidates are pruned using both an absolute threshold \(\tau\) and a relative peak threshold \(\tau_{\mathrm{rel}}\). Crucially, harmonization is triggered only when the count of qualified support views satisfies \(|\mathcal{S}_t^{\mathrm{sup}}| \ge K_{\min}\), ensuring no ambiguous pseudo-targets are generated under weak evidence.

2. 3D-Aware Multi-View Harmonizer with Mask-Aware Fast-Weight Updates: Repurposing NVS Priors for Inpainting

Off-the-shelf 2D generative diffusion models lack explicit 3D projective geometry awareness, yielding multi-view texture conflicts when used to fill occlusions. PriSplat repurposes a large-scale transformer-based novel view synthesis prior as a multi-view harmonizer. Because the NVS architecture naturally learns cross-view patch correspondence and view synthesis conditioned on rigid camera poses, providing unoccluded support views allows it to inpaint the occluded background while rigorously preserving 3D consistency.

However, the foundation NVS architecture updates internal scene-specific parameters (fast weights \(W\)) at test time. If candidate support views contain residual distractor artifacts, dynamic noise leaks directly into the scene memory. To eliminate distractor leakage, PriSplat incorporates a mask-aware fast-weight update. It computes token-level query-key alignment \(\alpha = \langle Q, K \rangle_{\mathrm{token}}\) and generates an adaptive down-weighting gate \(b = \sqrt{\sigma(\alpha)}\) to modulate token updates:

\[\Delta W = \texttt{Update}(W; b \odot K, b \odot V, \eta), \qquad W' \leftarrow W + \Delta W\]

By suppressing weakly aligned or transient tokens, \(W'\) encodes only static 3D background evidence. Querying target pose tokens \(Q_t\) then synthesizes a clean, perspective-accurate harmonized view \(\hat{I}_t = \texttt{Harmonizer}(\{(\tilde{I}_j, P_j)\}, P_t)\) that serves as reliable pseudo-ground truth.

3. Three-Stage Progressive Hybrid Supervision Schedule: Balancing Geometric Scaffolding and High-Frequency Detail

Rather than continuously applying the large harmonizer throughout the entire 30k training iterations, PriSplat uses an efficient phased training curriculum. In Stage 1 (0 to 10k iterations), standard masked reconstruction loss \(\mathcal{L}_{\mathrm{mask}}\) optimizes 3DGS on uncorrupted pixels, establishing a sufficiently reliable geometric scaffold to accurately compute Fisher projection intersections.

In Stage 2 (10k to 20k iterations), multi-view harmonization is enabled. For target views with sufficient support evidence (\(|\mathcal{S}_i^{\mathrm{sup}}| \ge K_{\min}\)), dense full-image supervision using L1 and D-SSIM is applied against the harmonized pseudo-ground truth:

\[\mathcal{L}_i^{\mathrm{stage2}} = \sum_{p \in \Omega} \mathcal{D}(f_{p,i}(\theta), \hat{I}_i(p))\]

Views lacking sufficient support fall back safely to standard masked loss \(\mathcal{L}_{\mathrm{mask}}\). In Stage 3 (20k to 30k iterations), the pipeline turns off the harmonizer and switches back to masked reconstruction loss on original static pixels. This final stage allows the scene to refine fine-grained textures and sharpness directly from uncorrupted real observations, eliminating any minor blur or oversmoothing introduced by generative completion.

Key Experimental Results

Main Results

Quantitative comparisons were conducted on the 6 standard benchmark scenes from the real-world On-the-go dataset across varying occlusion levels: low occlusion (Mountain, Fountain), medium occlusion (Corner, Patio), and high occlusion (Spot, Patio-high). The evaluation metrics include PSNR↑, SSIM↑, and LPIPS↓. Baselines include SLS-mlp, WildGaussians, OCSplats, ForestSplats, HybridGS, DeSplat, DeGauss, and RobustSplat.

Occlusion Level	Scene Name	Metric	PriSplat (Ours)	RobustSplat [16]	DeSplat [61]	Gain (vs. Best Baseline)
Low Occlusion	Mountain	PSNR / SSIM / LPIPS	21.65 / 0.7080 / 0.1688	21.42 / 0.6934 / 0.1761	19.92 / 0.6696 / 0.1780	+0.23 dB / -0.0073 LPIPS
Low Occlusion	Fountain	PSNR / SSIM / LPIPS	21.42 / 0.7066 / 0.1590	20.97 / 0.6967 / 0.1603	20.66 / 0.6830 / 0.1673	+0.45 dB / +0.0099 SSIM
Medium Occlusion	Corner	PSNR / SSIM / LPIPS	25.98 / 0.8569 / 0.0992	25.76 / 0.8513 / 0.1077	24.61 / 0.8178 / 0.1312	+0.22 dB / -0.0085 LPIPS
Medium Occlusion	Patio	PSNR / SSIM / LPIPS	21.97 / 0.8215 / 0.1052	21.94 / 0.8182 / 0.1054	21.63 / 0.8122 / 0.1093	+0.03 dB / +0.0033 SSIM
High Occlusion	Spot	PSNR / SSIM / LPIPS	25.15 / 0.8070 / 0.1149	24.95 / 0.8035 / 0.1179	24.53 / 0.7830 / 0.1521	+0.20 dB / -0.0030 LPIPS
High Occlusion	Patio-high	PSNR / SSIM / LPIPS	22.48 / 0.7559 / 0.1593	22.08 / 0.7538 / 0.1622	21.34 / 0.7377 / 0.1880	+0.40 dB / +0.0021 SSIM
Overall Mean	Average across 6 Scenes	PSNR / SSIM / LPIPS	23.11 / 0.7760 / 0.1344	22.86 / 0.7710 / 0.1387	22.28 / 0.7589 / 0.1543	+0.25 dB / -0.0043 LPIPS

(Extracted from Table 1 of the original paper; DeGauss obtains 21.71 dB on Mountain, but lags behind on Patio-high and in overall mean metrics.)

Ablation Study

The ablation experiments systematically isolate the contributions of multi-view harmonization, geometry-aware selection, threshold filtering, continuous mask learning, and mask-aware fast-weight updates (mean metrics from Table 2 of the paper):

Configuration Variant	Harmonization	Threshold	Support Selection	Mask Update	Mask-aware Update	PSNR↑	SSIM↑	LPIPS↓	Configuration Note
(a) Mask-only Baseline	✗	✗	✗	✓	✗	22.86	0.7710	0.1387	Baseline setting suppressing distractor loss to zero
(b) Naive KNN	✓	✗	KNN	✓	✓	22.72	0.7673	0.1415	Inappropriate views inject severe completion errors
(c) Geometry-aware w/o Thresh	✓	✗	Geometry-aware	✓	✓	22.85	0.7711	0.1375	Unreliable views harm optimization without gating
(d) KNN with Threshold	✓	✓	KNN	✓	✓	22.89	0.7706	0.1386	Gating helps, but spatial visibility modeling is missing
(e) Frozen Mask Predictor	✓	✓	Geometry-aware	✗	✓	22.89	0.7712	0.1404	Initial mask errors remain uncorrected
(f) Standard Fast-weight	✓	✓	Geometry-aware	✓	✗	22.93	0.7736	0.1373	Without token gating, transient cues pollute memory
(g) Full Model (PriSplat)	✓	✓	Geometry-aware	✓	✓	23.11	0.7760	0.1344	Synergistic combination achieving best overall metrics

In addition, comparison across inpainting priors (Table 5 of the paper) reveals that LaMa-based 3DGIC (22.26 dB PSNR) and FLUX-based RogSplat (21.62 dB PSNR) underperform PriSplat (23.11 dB PSNR), demonstrating that scale alone without 3D geometric consistency is insufficient.

Key Findings

• Geometry-aware support selection is critical for stabilization: Merely activating harmonization with standard KNN view selection (variant b, 22.72 dB) degrades performance below the mask-only baseline (variant a, 22.86 dB). This demonstrates that feeding poorly positioned or occluded support views to generative priors produces cross-view inconsistencies that derail 3DGS convergence.
• Dramatic improvements in heavily occluded scenes: In Patio-high, where severe pedestrian occlusions leave massive holes, PriSplat achieves 22.48 dB PSNR (+0.40 dB over RobustSplat). Qualitative renderings demonstrate that while baselines produce spiky floating Gaussians and blurry ground surfaces, PriSplat restores crisp, planar stone paver textures.
• Intermediate injection scheduling prevents generative drift: Hyperparameter analysis (Table 3a) confirms that restricting harmonized supervision to the middle iterations (10k to 20k) achieves the highest PSNR (23.11 dB) compared to extending it to 30k (22.92 dB). This shows that generative priors should guide structure completion during early convergence, leaving final iterations to lock onto real photographic ground truth.

Highlights & Insights

• NVS Models as 3D-Aware Inpainting Engines: Rather than attempting to force 2D diffusion models into multi-view tasks via heavy regularizers, PriSplat astutely frames novel view synthesis as pose-conditioned reference inpainting, exploiting transformer-based NVS models to inherently respect rigid 3D spatial transformations.
• Gradient-Based Fisher Information for Occlusion Sensitivity: Computing the Frobenius norm of back-propagated color gradients provides an elegant, principled proxy for measuring whether a background 3D Gaussian is reliably observable from a candidate view, turning qualitative visibility into an explicit numeric density metric.
• Minimal Practical Training Overhead: By triggering the harmonizer only during an intermediate training window for qualifying views (4.36 s per target view at 512×512 resolution), PriSplat requires only 7.3 GB peak VRAM and 0.38 hours total training time on Patio-high, contrasting sharply with complex dynamic-static decoupling methods like DeGauss (17.1 GB average VRAM, 7.41 hours).

Limitations & Future Work

• Static Illumination and Reflectance Assumption: The current harmonizer formulation assumes consistent Lambertian lighting across support and target views. In long-duration outdoor captures with rapid sun movement, cloud shadows, or specular highlights, harmonized pseudo-ground truth may exhibit color cast discrepancies.
• Degradation in Fully Occluded Blind Spots: If an occluded background region is never visible in any training camera view (\(|\mathcal{S}_t^{\mathrm{sup}}| < K_{\min}\)), the algorithm conservatively falls back to standard zero-masking, leaving the unobserved area uncompleted.
• Future Directions: Integrating explicit per-view appearance embeddings or exposure adaptation into the multi-view harmonizer would allow robust handling of extreme lighting shifts; extending the pipeline to feed forward 3D Gaussian diffusion priors represents another promising avenue.

Related Work & Insights

• vs. RobustSplat [16] / SpotLessSplats [48]: Standard distractor-free 3DGS baselines rely entirely on identifying and zero-masking distractors, which abandons supervision over occluded regions and inevitably spawns floating Gaussians in novel views. PriSplat retains robust mask estimation while actively re-establishing dense 3D-consistent supervision via multi-view harmonization.
• vs. RogSplat [25] / 3DGIC [18]: Prior generative inpainting methods leverage 2D diffusion or Fourier convolution models (such as FLUX or LaMa) that synthesize textures on independent image planes, causing severe cross-view texture shimmering and geometric tearing. PriSplat employs pose-conditioned NVS priors and mask-aware fast-weight updates, ensuring strict 3D geometric alignment across synthesized views.

Rating

• Novelty: ⭐⭐⭐⭐☆ (Reframes distractor-free 3DGS from naive loss masking to 3D-aware NVS prior harmonization)
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ (Exhaustive benchmarking across On-the-go and RobustNeRF datasets with rigorous component ablations and prior comparisons)
• Writing Quality: ⭐⭐⭐⭐⭐ (Flawless logical progression, crisp mathematical formulations, and insightful qualitative analyses)
• Value: ⭐⭐⭐⭐☆ (Provides an effective, computationally practical blueprint for clean 3DGS reconstruction in chaotic real-world environments)

Related Papers

• [ECCV 2026] DF3DV-1K: A Large-Scale Dataset and Benchmark for Distractor-Free Novel View Synthesis
• [ICLR 2026] Distractor-free Generalizable 3D Gaussian Splatting
• [CVPR 2025] DeSplat: Decomposed Gaussian Splatting for Distractor-Free Rendering
• [CVPR 2026] MU-GeNeRF: Multi-view Uncertainty-guided Generalizable Neural Radiance Fields for Distractor-aware Scene
• [ECCV 2026] Improving Sparse-View 3DGS Generalization via Flat Minima Optimization