SG-Splat: Stable Geometry Learning under Pose Uncertainty for Pose-Free 3D Gaussian Splatting

Xiaohan Lu; Jiaqi Zhang; Weijia Jiang; Siwei Ma

cs.CVOct 26, 2026

SG-Splat: Stable Geometry Learning under Pose Uncertainty for Pose-Free 3D Gaussian Splatting

Authors: Xiaohan Lu, Jiaqi Zhang, Weijia Jiang, Siwei Ma

Organizations: State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University, Beijing 100871, China · Migu Interactive Entertainment Co., Ltd, Nanjing 210000, China

Abstract

3D Gaussian Splatting (3DGS) has become a promising representation for real-time novel view synthesis and 3D scene reconstruction. For practical deployment, recent studies have explored pose-free feed-forward 3DGS, which reconstructs Gaussian representations directly from unposed input images. However, these methods remain sensitive to residual pose errors, which perturb canonical-space Gaussian transformation, introducing noisy supervision for geometry and attribute learning. To address this problem, we present SG-Splat, a pose-free feed-forward 3DGS framework that improves robustness to pose uncertainty and stabilizes the learning of Gaussian primitives. SG-Splat decouples geometry estimation and Gaussian attribute prediction into two dedicated branches, reducing error coupling between pose-sensitive geometry and Gaussian attributes. We further introduce a coarse-to-fine refinement pipeline, consisting of Flow-Guided Pose Refinement (FGPR) and Residual-Guided Geometry Correction (RGC), to improve geometric alignment before Gaussian attribute reconstruction. Specifically, FGPR refines camera poses with flow-based correlation features and differentiable bundle adjustment. RGC then corrects initial point maps using depth residuals derived from FGPR, producing more reliable Gaussian centers. To stabilize training, we introduce a dual-path strategy that mixes ground-truth and predicted poses for canonical transformation. It provides clean supervision for Gaussian attributes while improving geometry robustness to residual pose errors. Extensive experiments on DL3DV and RealEstate10K show that SG-Splat achieves state-of-the-art performance in both novel view synthesis and pose estimation.

Figures & tables

Fig. 1: Comparison of pose usage strategies in pose-free feed-forward 3DGS training. (a) AnySplat [ 1 ] uses predicted poses with pose supervision, maintaining consistency between training and inference but suffering from perturbations caused by pose errors. (b) WorldMirror [ 2 ] uses ground-truth poses with pose supervision, stabilizing training but introducing inconsistency between training and inference. (c) E-RayZer [ 3 ] relies on predicted poses without pose supervision, making optimization sensitive to pose errors. (d) Our method uses mixed poses with pose supervision, balancing stable learning and robustness to residual pose errors.

Fig. 2: Overview of the pipeline. The input images are processed by separate geometry and attribute branches. The geometry branch predicts initial local point maps and camera poses through the point head and camera head, while the attribute branch predicts the remaining Gaussian attributes. Flow-Guided Pose Refinement (FGPR) refines the initial poses and produces depth residuals, which are used by Residual-Guided Geometry Correction (RGC) to correct the local Gaussian centers $\boldsymbol{\mu}^{\mathrm{local}}$ . The corrected centers are transformed into the canonical space with the refined poses and combined with the predicted attributes to form the final Gaussian representation.

Fig. 3: Flow-Guided Pose Refinement (FGPR). At each iteration, FGPR uses the current pose and geometry estimates to build flow-based correlation features and optical flow of camera motion. Together with context features, these cues are fed into an update operator to predict correspondence corrections $\Delta\mathbf{p}_{ij}$ . The corrected correspondences $\mathbf{p}^{*}_{ij}$ are optimized by differentiable bundle adjustment (DBA) to jointly update camera poses and depths. Over $T$ iterations, FGPR progressively improves pose accuracy and produces depth residuals for subsequent geometry correction.

Fig. 4: Residual-Guided Geometry Correction (RGC). RGC uses the residual depth map

\Delta\hat{\mathbf{D}}_{i}

from FGPR to correct the initial local point map

\hat{\mathbf{M}}_{i}

. The residual depth map and the initial point map are embedded and fused with geometry tokens to predict a bounded multiplicative correction, producing the corrected point map

\tilde{\mathbf{M}}_{i}

Fig. 5: Dual-Path Training Strategy. The refined predicted pose and the ground-truth (GT) pose are sampled with equal probability. The selected pose is used to transform local Gaussian centers into the canonical space for rendering. When the ground-truth pose is selected, reliable rendering supervision updates both the geometry and attribute branches. When the predicted pose is selected, rendering gradients are detached from the attribute branch and only the optimize geometry branch. This asymmetric optimization stabilizes attribute learning while improving geometry robustness under residual pose errors.

Method	GT Pose	12v			18v			24v			32v
Method	GT Pose	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
MVSplat	✓	21.149	0.696	0.236	19.452	0.632	0.281	19.761	0.641	0.272	17.912	0.492	0.403
DepthSplat	✓	22.027	0.763	0.206	19.712	0.673	0.294	20.110	0.695	0.275	18.240	0.601	0.365
YoNoSplat	✓	23.285	0.773	0.177	22.428	0.747	0.199	22.664	0.758	0.192	21.728	0.727	0.217
Ours	✓	23.487	0.781	0.168	22.615	0.764	0.179	23.163	0.775	0.174	22.255	0.745	0.187
NoPoSplat		19.182	0.543	0.338	17.325	0.501	0.401	17.641	0.498	0.399	17.024	0.481	0.410

TABLE I: Quantitative comparisons on DL3DV [ 57 ] under different input-view settings. GT Pose indicates whether camera poses are provided. Our method consistently achieves the best performance from 12 to 32 input views in both posed and pose-free settings.

Fig. 6: Qualitative comparisons on the DL3DV dataset with 24 input views. With more accurate camera pose estimation, our model produces geometrically consistent Gaussian primitives, yielding sharper boundaries, finer details, and fewer artifacts in novel views.

Method	GT Pose	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
MVSplat	✓	22.456	0.761	0.223
DepthSplat	✓	24.172	0.848	0.141
Ours	✓	25.041	0.849	0.132
NoPoSplat		22.044	0.731	0.210
AnySplat		23.712	0.748	0.186
WorldMirror		23.795	0.844	0.152

TABLE II: Quantitative comparisons on RealEstate10K [ 59 ] under the 6-view setting. GT Pose indicates whether camera poses are provided.

Fig. 7: Qualitative comparisons on the RealEstate10K dataset with sparse-view inputs (6 views). Even with only a few input views, our model reconstructs more coherent scene geometry and more realistic novel views, with clearer details and reduced artifacts.

Method	DL3DV												RealEstate10K
	12v			18v			24v			32v			6v
	$5^{\circ}$	$10^{\circ}$	$20^{\circ}$	$5^{\circ}$	$10^{\circ}$	$20^{\circ}$	$5^{\circ}$	$10^{\circ}$	$20^{\circ}$	$5^{\circ}$	$10^{\circ}$	$20^{\circ}$	$5^{\circ}$	$10^{\circ}$	$20^{\circ}$
MASt3R	0.626	0.773	0.874	0.599	0.756	0.866	0.638	0.784	0.883	0.653	0.788	0.881	0.521	0.727	0.770
VGGT	0.680	0.838	0.919	0.654	0.821	0.908	0.653	0.824	0.911	0.629	0.809	0.903	0.590	0.763	0.870
${\pi^{3}}$	0.837	0.918	0.959	0.792	0.895	0.947	0.829	0.914	0.957	0.797	0.899	0.949	0.725	0.852	0.921
NoPoSplat	0.412	0.625	0.744	0.393	0.612	0.735	0.401	0.635	0.751	0.390	0.625	0.739	0.431	0.619	0.749

TABLE III: Pose estimation results on DL3DV and RealEstate10K. The upper block reports 3D foundation models, and the lower block reports pose-free feed-forward 3DGS methods. Our method achieves the best AUC across dense-view DL3DV settings and sparse-view RealEstate10K.

Variant	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Base	22.361	0.735	0.203
FGPR*	22.499	0.740	0.201
FGPR + Pred Pose	21.155	0.647	0.213
FGPR + GT Pose	22.912	0.751	0.192
FGPR + Mix Pose	23.078	0.763	0.181
FGPR + Mix Pose + RGC	23.190	0.772	0.171

TABLE IV: Ablation study on DL3DV with 24 input views. FGPR* indicates that FGPR is trained, with all others frozen. The base model adopts the same architecture but excludes both FGPR and RGC, and is trained with ground-truth (GT) poses.

$T$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	AUC@5 ${}^{\circ}\uparrow$
1	23.079	0.761	0.179	0.794
2	23.190	0.772	0.171	0.849
3	23.183	0.770	0.170	0.851

TABLE V: Effect of the number of FGPR iterations

T

on DL3DV with 24 input views.

3D Gaussian Splatting

3D Gaussian Splatting represents scenes as optimizable 3D Gaussian primitives for fast, photorealistic rendering from new viewpoints.

See all 1,013 papers

SPFSplatV2: Efficient Self-Supervised Pose-Free 3D Gaussian Splatting from Sparse Views
Sep 21, 2025Ranran Huang, Krystian Mikolajczyk3D Gaussian Splatting 3D Reconstruction
Department of Electrical and Electronic Engineering, Imperial College London, London SW7 2AZ, United Kingdom
DeltaSplat: Iterative Gaussian Refinement for Pose-Free Feed-Forward 3D Gaussian Splatting
Oct 7, 2026Chanung Park, Seunghyeon Song, Joo Chan Lee +23D Gaussian Splatting 3D Reconstruction
Department of Electrical and Computer Engineering, Sungkyunkwan University · Department of Artificial Intelligence, Sungkyunkwan University · Department of Artificial Intelligence, Yonsei University
G3Splat: Geometrically Consistent Generalizable Gaussian Splatting
Dec 19, 2025Mehdi Hosseinzadeh, Shin-Fang Chng, Yi Xu +33D Gaussian Splatting 3D Reconstruction
Australian Institute for Machine Learning · Goertek Alpha Labs · MBZUAI

Sep 21, 2025J/K move · Enter open · S save

SPFSplatV2: Efficient Self-Supervised Pose-Free 3D Gaussian Splatting from Sparse Views

Ranran Huang, Krystian Mikolajczyk

Department of Electrical and Electronic Engineering, Imperial College London, London SW7 2AZ, United Kingdom

3D Gaussian Splatting 3D Reconstruction Camera Pose Estimation Novel View Synthesis Sparse-View 3D Reconstruction Feed-Forward 3D Reconstruction

Open paper AskPDF arXiv

We introduce SPFSplatV2, an efficient feed-forward framework for 3D Gaussian splatting from sparse multi-view images, requiring no ground-truth poses during training or inference. The framework employs a shared feature extraction backbone to jointly predict 3D Gaussian primitives and camera poses in a canonical space from unposed inputs. To enable efficient and accurate pose estimation, we introduce a masked attention mechanism for target-view pose prediction and a reprojection loss that enforces pixel-aligned Gaussian primitives, providing stronger geometric constraints. We further demonstrate the compatibility of our training framework with different reconstruction architectures, resulting in two model variants. Remarkably, despite the absence of pose supervision, our method achieves state-of-the-art performance in both in-domain and out-of-domain novel view synthesis, even under extreme viewpoint changes and limited image overlap. It also surpasses many methods that rely on geometric supervision in relative pose estimation. By eliminating dependence on ground-truth poses, our method offers the scalability to leverage larger and more diverse datasets. Code and pretrained models will be available on our project page: https://ranrhuang.github.io/spfsplatv2/.

Abstract

3D Gaussian Splatting

Explore similar work

SPFSplatV2: Efficient Self-Supervised Pose-Free 3D Gaussian Splatting from Sparse Views

DeltaSplat: Iterative Gaussian Refinement for Pose-Free Feed-Forward 3D Gaussian Splatting

G3Splat: Geometrically Consistent Generalizable Gaussian Splatting