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

Yejun Zhang; Zihan Wang; Xu Ji; Yihao Wang; Yuxin Hou; Junyuan Fang; Juho-Matti Kilpeläinen; Arno Solin

cs.CVAug 17, 2026

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

Authors: Yejun Zhang, Zihan Wang, Xu Ji, Yihao Wang, Yuxin Hou, Junyuan Fang, Juho-Matti Kilpeläinen, Arno Solin, +3 more

Organizations: Aalto University, Finland · Deep Render, UK · ELLIS Institute Finland · Nokia Technologies, Finland · Tampere University, Finland · University of Oulu, Finland

Abstract

Generating photorealistic novel views from unposed images requires both 3D geometric understanding and the ability to synthesize unseen content. A natural strategy combines feed-forward 3DGS reconstruction with multi-view diffusion. Yet prior pipelines extract at most one signal from the reconstruction, either pixel rendering or learned features, while none exploits per-Gaussian visibility for occlusion-aware reference selection. This information disconnect leaves renderable geometry, visibility cues, and learned features unused. SplatGuide closes this disconnect by reusing a single 3DGS scene across three complementary roles. Rendered images provide pixel-aligned geometric conditioning. Per-Gaussian source-view indices are rendered into a target-view voting map for occlusion-aware reference selection. Reconstruction tokens supply feature-level guidance via cross-attention. All three signals derive from the same reconstruction forward pass. Across RealEstate10K, DL3DV, Tanks-and-Temples, and Mip-NeRF 360, SplatGuide achieves state-of-the-art pose-free novel view synthesis. On RealEstate10K, with a moderate number of input views, it surpasses the ground-truth-pose baseline.

Figures & tables

Figure 2: Overview of SplatGuide. Reconstruction: A feed-forward model takes $N$ unposed images and produces 3D Gaussians with per-Gaussian source-view indices $(\mu,\sigma,r,s,c,\mathit{index})$ , along with camera and register tokens. View Selection: The per-Gaussian indices are rendered into a target-view index map; pixel-wise voting retrieves the top- $K$ reference views, whose RGBs and Plücker coordinates form the selected context views. Generation: The rendered target RGB, noise, and target Plücker are concatenated with the context views and fed into the diffusion model, which is further conditioned on reconstruction tokens via cross-attention to produce the final photorealistic image.

Figure 3: Overview of the proposed view selection pipeline. We reuse the reconstructed 3D Gaussians with per-Gaussian source-view indices, render an occlusion-aware target-view index map via first-hit visibility, and aggregate pixel votes into per-view visibility scores $S(k)$ to determine Top-K Indices. The final selection is refined by DeDup, which filters spatially redundant candidates, and PoseAug, which augments the context set with pose-proximal views for isolated targets, ensuring a complete and diverse set of $K$ reference images.

Panel of Figure 4 — Figure 4: Visual comparison of novel view synthesis. Columns from left to right: ground truth, the coarse 3DGS render serving as our geometric prior, our SplatGuide result, SEVA, and ViewCrafter. SplatGuide produces sharper structures and fewer artifacts than both baselines, benefiting from the explicit geometric guidance provided by the rendered image and injected tokens.

RealEstate10K
	GT	3-view			6-view			9-view
Method	pose	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
SEVA	✓	27.57	0.89	0.07	29.24	0.90	0.06	29.63	0.91	0.05
AnySplat	×	19.81	0.71	0.24	23.23	0.79	0.16	24.13	0.82	0.15
WorldMirror	×	21.13	0.77	0.16	22.89	0.81	0.12	23.35	0.83	0.11
Fillerbuster	×	19.13	0.63	0.27	19.81	0.65	0.24	20.59	0.67	0.22

Table 1: Quantitative comparison on the in-domain RealEstate10K (top) and DL3DV (bottom) datasets. Best results among unposed methods are in bold , and second-best are underlined .

Regression-based methods:
	Tanks and Temples						Mip-NeRF 360
Method	3-view			6-view			3-view			6-view			9-view
	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
AnySplat	16.04	0.46	0.35	18.00	0.52	0.30	8.26	0.20	0.66	10.30	0.23	0.60	11.42	0.26	0.56
WorldMirror	17.45	0.55	0.27	19.06	0.56	0.25	11.90	0.24	0.61	12.52	0.26	0.55	13.07	0.26	0.52
Diffusion-based methods:

Table 2: Quantitative comparison on out-of-domain datasets. All methods operate without ground-truth poses. Best results are in bold , and second-best are underlined .

Panel of Figure 5 — Figure 5: View extrapolation under large viewpoint changes. For each scene we show (a) the selected context views, (b) coarse 3DGS renders at target poses far from all context cameras, and (c) the corresponding generated novel views. The degraded renders still anchor the visible layout, which the diffusion model completes into coherent images.

	$B{=}6$ (sparse)			$B{=}9$ (mid)			$B{=}16$ (ample)
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Temporal(DFoT)	20.88	0.71	0.22	21.50	0.75	0.18	26.53	0.86	0.10
CamDist(SEVA)	24.34	0.80	0.11	25.44	0.81	0.10	27.71	0.88	0.07
FoV(WorldMem)	26.52	0.83	0.09	27.61	0.84	0.07	29.25	0.90	0.06
Surfel(Vmem)	26.93	0.84	0.09	27.86	0.85	0.08	29.12	0.90	0.07
Ours	28.25	0.85	0.06	28.91	0.86	0.06	29.94	0.91	0.05

Table 3: Comparison of view selection policies with a fixed generator. All methods use the same reconstruction and diffusion backbones; only the selection policy differs.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Baseline + predicted pose	15.64	0.33	0.38
+ rendering image	16.30	0.35	0.33
+ rendering image + cam & reg	16.63	0.36	0.32

Table 4: Ablation study on geometric conditioning on DL3DV with 9 views. We progressively add rendering images, camera tokens and register tokens to the baseline to measure each component’s contribution.

Panel of Figure 6 — Figure 6: Qualitative comparison of view-selection policies. Top two rows: our method recovers fine details by selecting complementary views. Bottom two rows: our visibility-aware selection suppresses occluder-dominated candidates and reveals hidden geometry.

Pool size $V$	32	64	96	128
PSNR $\uparrow$	16.06	16.40	16.87	17.21
SSIM $\uparrow$	0.41	0.42	0.43	0.47
LPIPS $\downarrow$	0.28	0.25	0.25	0.23

Table 5: Scaling the candidate pool on long-trajectory Tanks-and-Temples scenes (Long-LRM split) with a fixed context budget

B{=}9

. Larger pools consistently improve quality without degrading palette-encoded index recovery.

Panel of Figure 7 — Figure 7: Novel view synthesis on large-scale scenes with 100+ input images. Each group shows four example inputs (left $2{\times}2$ grid) and two generated novel views (right), on Nerfbusters [ 29 ] , Tanks-and-Temples, and self-collected casual captures.

	3-view			6-view			9-view
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Ours (AnySplat)	26.56	0.83	0.08	29.97	0.87	0.05	31.29	0.88	0.04
Ours (WorldMirror)	26.52	0.84	0.07	29.29	0.87	0.04	30.00	0.88	0.04

Table 6: Zero-shot generalization across reconstruction backbones on RealEstate10K. Our diffusion model, trained exclusively with WorldMirror priors, successfully transfers to AnySplat without retraining.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Method	${\bf R}_{err}$ ( ∘ ) $\downarrow$	${\bf T}_{err}$ $\downarrow$
DUSt3R	0.69	0.06
WorldMirror	0.27	0.02

Appendix

Table A1: Camera pose error comparison between DUSt3R and WorldMirror as reconstruction backbones.

Panel of Figure 14 — Figure A1 : Qualitative Results on the MipNeRF360, Tanks and Temples, and DL3DV Datasets. Red and blue boxes highlight detailed regions with corresponding magnified views showing texture fidelity and geometric accuracy.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
CamDist	18.37	0.58	0.31
Surfel	18.52	0.59	0.30
FoV	18.27	0.58	0.30
Ours	19.30	0.61	0.26

Appendix

Table A2: Quantitative comparison of view-selection policies on our captured real-world scenes with a 6-view context budget. Metrics are averaged over all target views.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Ours	28.25	0.85	0.06
w/o PoseAug	26.50	0.83	0.08
w/o DeDup	27.94	0.84	0.06
w/o PoseAug + DeDup	25.13	0.81	0.11

Appendix

Table A3: Ablation of selection variants on RealEstate10K with 32 candidates and a 6-view context budget.

Stage	Component	Time (s)
Reconstruction	WorldMirror forward pass	3.73
Selection	3DGS rendering (RGB + index map)	0.005
	full selection incl. DeDup and PoseAug	0.006
Generation	SEVA sampling (baseline)	62.9
	SplatGuide sampling (ours)	71.8
End-to-end (ours)		75.5

Appendix

Table A4: Inference cost on a single RTX 4090 with batch size 1 and a candidate pool of 32 views. Sampling times for SEVA and for our model are measured under the same DDIM configuration. Rendering and selection together account for less than

0.01\%

of our end-to-end latency. The reconstruction cost is not exclusive to our method: the unposed SEVA baseline also requires a reconstruction pass to obtain poses, so the two pipelines differ mainly in sampling time.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
ViewCrafter (SEVA)	22.81	0.83	0.16
ViewCrafter (All-view)	22.80	0.81	0.16
ViewCrafter (Test-align)	20.18	0.74	0.23
Ours	26.52	0.84	0.07

Appendix

Table A5: Comparison of ViewCrafter results under different evaluation protocols on RealEstate10K benchmark. Here we use 3 input-views as an example.

Explore similar work

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
GAPS: Generative Active Pseudo-view Selection for Sparse-View 3D Gaussian Splatting
Sep 20, 2026Hongfei Zhu, Haochen Deng, Sitao Zhang +13D Gaussian Splatting Active Perception
Shanghai Jiao Tong University, 800 Dongchuan Road, Shanghai, 200240, Shanghai, China · Goertek Inc., No. 268 Dongfang Road, Hi-tech Industrial Development Zone, Weifang, 261031, Shandong, China · Fudan University, 220 Handan Road, Shanghai, 200433, Shanghai, China
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

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

Explore similar work

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

GAPS: Generative Active Pseudo-view Selection for Sparse-View 3D Gaussian Splatting

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