GeoGSplat: Geometry-Grounded Non-Pixel-Aligned Gaussian Splatting

Jiale Wang; Chen Zhao; Haoyang Zhang; Yanhao Wu; Fei He; Yanhu Shan; Rui Wu; Wei Ke

cs.CVOct 27, 2026

GeoGSplat: Geometry-Grounded Non-Pixel-Aligned Gaussian Splatting

Authors: Jiale Wang, Chen Zhao, Haoyang Zhang, Yanhao Wu, Fei He, Yanhu Shan, Rui Wu, Wei Ke,

Organizations: Xi’an Jiaotong University · Horizon Robotics · EPFL · University of the Chinese Academy of Sciences

Abstract

Feed-forward Gaussian Splatting enables efficient novel view synthesis, but pixel-aligned prediction makes the primitive count grow with image resolution and view count, leading to substantial redundancy and increased computational cost. Non-pixel-aligned methods address this limitation by predicting a fixed number of Gaussians. Without explicit ray and depth constraints, these methods yield misplaced Gaussian centers, causing rendering artifacts under large viewpoint changes. To address this issue, we introduce GeoGSplat, a geometry-grounded feed-forward model that explicitly incorporates geometric priors into the non-pixel-aligned framework. A dual-branch encoder constructs Anchors and Triplane to capture coarse 3D structure and fine-grained scene features, respectively. Subsequently, a hierarchical decoder progressively decodes dense 3D points and Gaussian primitives, enabling 3D reconstruction and novel view synthesis. GeoGSplat surpasses both state-of-the-art pixel-aligned and non-pixel-aligned Gaussian Splatting methods on DTU and NRGBD in terms of novel view synthesis and reconstruction. The advantage of our geometry-grounded paradigm further increases as target viewpoints move farther from the input cameras, demonstrating stronger generalization to large viewpoint changes. Project Page: https://comedy0913.github.io/projects/geogsplat

Figures & tables

Figure 1: GeoGSplat introduces a geometry-grounded non-pixel-aligned Gaussian Splatting paradigm. It achieves high-quality 3D reconstruction and novel view synthesis by incorporating reliable geometric priors through complementary Anchor and Triplane representations. Larger bubble size indicates a larger number of Gaussians.

Figure 2: Comparison of pixel-aligned, non-pixel-aligned, and our GeoGSplat paradigms . Pixel-aligned methods rely on explicit geometry priors but produce an increasing number of primitives (Prim.) with more input views. Non-pixel-aligned methods predict a fixed number of primitives but lack geometric priors. GeoGSplat combines fixed-size prediction with explicit geometric priors.

Figure 3: Pipeline of GeoGSplat. Image features are encoded in parallel into Anchors and Triplane . Anchors comprise sparse 3D spheres with learnable centers and adaptive radii, accounting for the coarse scene structure; Triplane consists of three orthogonal 2D feature maps that encode fine-grained high-dimensional scene features. Anchor-to-Point Decoding (blue) gathers Triplane features within each Anchor region to predict offsets and recover dense 3D points; Point-to-Gaussian Decoding (orange) re-queries the Triplane at these points to predict the corresponding Gaussian primitives.

Method	DTU				NRGBD
	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	NUM	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	NUM
MVSplat	14.44	0.520	0.401	158K	18.53	0.609	0.319	255K
DepthSplat	15.46	0.599	0.331	158K	22.98	0.792	0.170	255K
AnySplat	16.60	0.630	0.166	170K	20.12	0.690	0.138	643K
AnySplat ∗	17.54	0.700	0.148	166K	20.52	0.756	0.140	624K
VolSplat	15.86	0.602	0.306	152K	22.22	0.757	0.216	246K

Table 1: Novel view synthesis with four input views. NUM denotes the number of Gaussians used for rendering. The best PSNR, SSIM, and LPIPS are shown in bold.

Method	DTU			NRGBD
	Pred $\rightarrow$ GT $\downarrow$	GT $\rightarrow$ Pred $\downarrow$	CD $\downarrow$	Pred $\rightarrow$ GT $\downarrow$	GT $\rightarrow$ Pred $\downarrow$	CD $\downarrow$
MVSplat	0.271	0.061	0.166	0.091	0.080	0.085
DepthSplat	0.130	0.066	0.098	0.093	0.058	0.076
AnySplat	0.048	0.032	0.040	0.063	0.046	0.055
AnySplat ∗	0.033	0.016	0.025	0.046	0.038	0.042
VolSplat	0.091	0.035	0.063	0.061	0.059	0.060

Table 2: 3D reconstruction with four input views. Pred

\rightarrow

GT and GT

\rightarrow

Pred denote Chamfer distances from predicted points to ground-truth points and vice versa, respectively. CD indicates their average. The best results are shown in bold.

Figure 4: Qualitative novel view synthesis with four input views. Columns show ground truth, AnySplat ∗ , TokenGS, and GeoGSplat from left to right. Red boxes highlight blurred details in TokenGS and shifted or duplicated structures in AnySplat. In contrast, GeoGSplat preserves finer details and more accurate scene structures in these regions.

Figure 5: NVS under varying viewpoint changes on NRGBD. The plots show PSNR (left) and the performance advantage over TokenGS (right) at different target-context distances. The performance gap between GeoGSplat and TokenGS becomes larger as the target-context distance increases.

Figure 6: 3D reconstruction on NRGBD. From left to right: ground truth, TokenGS Gaussian centers, opacity-filtered TokenGS centers, points unprojected from depth rendered by TokenGS, and GeoGSplat dense points. TokenGS contains substantial geometric outliers that remain after opacity filtering, while its depth-based reconstruction is also displaced from the ground truth. GeoGSplat predicts points closer to the ground-truth geometry and provides more complete scene coverage.

Variant	DTU				NRGBD
	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	CD $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	CD $\downarrow$
w/o Anchor	18.67	0.700	0.188	0.031	22.46	0.782	0.185	0.030
w/o Triplane	18.63	0.663	0.253	0.031	22.47	0.733	0.229	0.031
Full	19.79	0.728	0.183	0.019	23.70	0.798	0.164	0.019
w/o Anchor-to-Point Decoding	18.32	0.653	0.276	0.032	21.81	0.703	0.276	0.041
w/o Point-to-Gaussian Decoding	18.71	0.720	0.186	0.031	22.35	0.768	0.189	0.031

Table 3: Ablation studies . Full uses both representations, both decodings, and both geometry losses. The best result in each group is shown in bold.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Variant	DTU				NRGBD
	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	CD $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	CD $\downarrow$
TokenGS-R	13.75	0.571	0.386	0.277	22.48	0.728	0.234	0.180
TokenGS-R + Geometry loss	14.41	0.608	0.304	0.044	24.17	0.797	0.158	0.029
GeoGSplat	20.48	0.771	0.137	0.018	24.78	0.847	0.120	0.017

Appendix

Table 4: Comparison with retrained TokenGS under four input views. Both TokenGS variants use the same setting as GeoGSplat and differ only in whether the geometry loss is applied.

Table 5: Evaluation under varying input view counts on DTU and NRGBD. All methods are tested with 2 and 12 input views without retraining. We report PSNR (P) for novel view synthesis, Chamfer distance (CD) for 3D reconstruction, and the number of predicted Gaussians (NUM).

	MVSplat	DepthSplat	AnySplat ∗	VolSplat	TokenGS	Ours
PSNR $\uparrow$	16.24	16.06	18.65	19.79	21.51	21.59
CD $\downarrow$	0.106	0.063	0.081	0.058	0.132	0.015

Appendix

Table 6: Comparison on Mip-NeRF 360 with 12 input views. The best results are shown in bold.

Dataset	Views	Dense Points CD $\downarrow$	Gaussian Centers CD $\downarrow$
DTU	2V	0.037148	0.037130
DTU	4V	0.017984	0.017879
DTU	12V	0.014296	0.014189
NRGBD	2V	0.023718	0.023623
NRGBD	4V	0.017406	0.017485
NRGBD	12V	0.015325	0.015472

Appendix

Table 7: Consistency between dense points and Gaussian centers. We report Chamfer distance computed on the dense points from Anchor-to-Point decoding and the final Gaussian centers after Point-to-Gaussian decoding.

Figure 7: Comparison of reconstruction and novel view synthesis.

Component	Ablation model	Main-table model
Input resolution	$256\!\times\!256$	$256\!\times\!256$
Patch size	$8\!\times\!8$	$8\!\times\!8$
Image encoder blocks	6	8
Shared Attention blocks	8	12
Triplane resolutions	$64\!\rightarrow\!128\!\rightarrow\!256$	$64\!\rightarrow\!128\!\rightarrow\!256$
Anchors $\times$ points per Anchor	$4096\!\times\!16$	$4096\!\times\!16$

Appendix

Table 8: Architecture comparison between the smaller model used for ablations and the model used in the main tables.

3D Gaussian Splatting

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

See all 1,019 papers

GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens
Apr 16, 2026Roni Itkin, Noam Issachar, Yehonatan Keypur +33D Gaussian Splatting 3DGS Acceleration
The Hebrew University of Jerusalem · Westlake University
VolSplat: Rethinking Feed-Forward 3D Gaussian Splatting with Voxel-Aligned Prediction
Sep 23, 2025Weijie Wang, Yeqing Chen, Zeyu Zhang +9Multi-View Consistency 3D Gaussian Splatting
Zhejiang University · GigaAI · The Chinese University of Hong Kong +4
SplatGuide: Geometric Priors from 3D Gaussians for Pose-Free Novel View Synthesis
Aug 17, 2026Yejun Zhang, Zihan Wang, Xu Ji +8Multi-View Diffusion Models Novel View Synthesis
Aalto University, Finland · Deep Render, UK · ELLIS Institute Finland +3

Apr 16, 2026J/K move · Enter open · S save

GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens

Roni Itkin, Noam Issachar, Yehonatan Keypur, Xingyu Chen+2

The Hebrew University of Jerusalem · Westlake University

3D Gaussian Splatting 3DGS Acceleration Novel View Synthesis 3D Scene Representation Feed-Forward 3D Reconstruction

Open paper AskPDF arXiv

The efficient spatial allocation of primitives serves as the foundation of 3D Gaussian Splatting, as it directly dictates the synergy between representation compactness, reconstruction speed, and rendering fidelity. Previous solutions, whether based on iterative optimization or feed-forward inference, suffer from significant trade-offs between these goals, mainly due to the reliance on local, heuristic-driven allocation strategies that lack global scene awareness. Specifically, current feed-forward methods are largely pixel-aligned or voxel-aligned. By unprojecting pixels into dense, view-aligned primitives, they bake redundancy into the 3D asset. As more input views are added, the representation size increases and global consistency becomes fragile. To this end, we introduce GlobalSplat, a framework built on the principle of align first, decode later. Our approach learns a compact, global, latent scene representation that encodes multi-view input and resolves cross-view correspondences before decoding any explicit 3D geometry. Crucially, this formulation enables compact, globally consistent reconstructions without relying on pretrained pixel-prediction backbones or reusing latent features from dense baselines. Utilizing a coarse-to-fine training curriculum that gradually increases decoded capacity, GlobalSplat natively prevents representation bloat. On RealEstate10K and ACID, our model achieves competitive novel-view synthesis performance while utilizing as few as 16K Gaussians, significantly less than required by dense pipelines, obtaining a light 4MB footprint. Further, GlobalSplat enables significantly faster inference than the baselines, operating under 78 milliseconds in a single forward pass. Project page is available at https://r-itk.github.io/globalsplat/

Abstract

3D Gaussian Splatting

Explore similar work

GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens

VolSplat: Rethinking Feed-Forward 3D Gaussian Splatting with Voxel-Aligned Prediction

SplatGuide: Geometric Priors from 3D Gaussians for Pose-Free Novel View Synthesis