InfiniSplat v2: Analyzing and Improving Feed-Forward 3DGS from Any Views

Yongzhen Hu; Hao Yu; Jiawei Wang; Xianming Zeng; Jianyun Xu; Xingliang Liu; Ruizhen Hu; Xiaowei Zhou

cs.CVOct 14, 2026

InfiniSplat v2: Analyzing and Improving Feed-Forward 3DGS from Any Views

Authors: Yongzhen Hu, Hao Yu, Jiawei Wang, Xianming Zeng, Jianyun Xu, Xingliang Liu, Ruizhen Hu, Xiaowei Zhou,

Organizations: Zhejiang University · Hello TransTech · Shenzhen University · Causa Robotics

Abstract

Feed-forward Gaussian reconstruction builds 3D scenes directly from input images for novel view rendering, but learning Gaussian attributes from predicted geometry for high-quality rendering under large viewpoint changes remains challenging. We analyze limitations in Gaussian attribute learning and introduce InfiniSplat v2, which improves rendering quality through architectural and training changes. Our method decouples geometry and attribute prediction with a frozen geometry branch, aligns predicted Gaussians using ground-truth camera poses during training, and adopts large-angle target view supervision. Across six datasets, the feed-forward outputs of the same trained model achieve state-of-the-art mean LPIPS and DISTS among the evaluated methods in both pose-conditioned and pose-free settings. In the pose-conditioned setting, mean LPIPS and DISTS decrease by 25.8% and 32.9%, respectively, compared with the strongest evaluated baselines. We further find that opacity learned with only image reconstruction supervision suppresses geometrically inconsistent points. Project page: https://zju3dv.github.io/InfiniSplat2/

Figures & tables

Figure 1: From point clouds to Gaussian scenes. InfiniSplat v2 retains point positions from a geometry foundation model and learns Gaussian attributes for high-quality novel view rendering. Left: input views, point clouds, and corresponding Gaussian renderings, with enlarged regions outlined in orange. Right: additional examples comparing point clouds with the resulting Gaussian scenes.

Figure 2: Training pipeline of InfiniSplat v2’s feed-forward reconstructor. A frozen geometry branch predicts depth and cameras, while a trainable branch predicts Gaussian attributes. Scale-aligned GT cameras construct the point cloud; after opacity pruning, input and large-angle target views supervise Gaussian learning.

Input	Method	DL3DV		ETH3D		NRGBD		Mip-NeRF 360		RE10K		DTU		Mean
		LPIPS $\downarrow$	DISTS $\downarrow$	LPIPS $\downarrow$	DISTS $\downarrow$	LPIPS $\downarrow$	DISTS $\downarrow$	LPIPS $\downarrow$	DISTS $\downarrow$	LPIPS $\downarrow$	DISTS $\downarrow$	LPIPS $\downarrow$	DISTS $\downarrow$	LPIPS $\downarrow$	DISTS $\downarrow$
Pose-free	AnySplat	0.5985	0.2490	0.5473	0.2297	0.1895	0.1131	0.5512	0.2259	0.2454	0.1113	0.4800	0.2345	0.4353	0.1939
	C3G	0.7503	0.3694	0.6722	0.3471	0.5393	0.2708	0.7390	0.3878	0.4600	0.2145	0.7115	0.3704	0.6454	0.3267
	QuerySplat	0.5601	0.2260	0.5805	0.2750	0.2586	0.1405	0.6076	0.2678	0.2158	0.1113	0.5199	0.2722	0.4571	0.2155
	DA3-GS	0.5146	0.2179	0.4950	0.2435	0.3456	0.1777	0.4738	0.2303	0.2526	0.1434	0.4108	0.2336	0.4154	0.2077
	WorldMirror 2.0	0.4800	0.2148	0.4917	0.2310	0.2099	0.1556	0.4701	0.2169	0.2210	0.1323	0.3812	0.2106	0.3757	0.1935

Table 1: Novel view synthesis on six datasets. Scores average

N=4,8,16

equally; Mean weights datasets equally. Ours/Ours ∗ : feed-forward/refined. First , second , and third are highlighted per input setting; best in bold.

Figure 3: Object-level qualitative comparisons under multi-view input settings.

Figure 4: Scene-level qualitative comparisons under multi-view input settings.

Figure 6: Effect of reconstruction components on surface artifacts and structural detail.

Figure 7: Opacity distributions and the effect of pruning on geometrically inconsistent points. GT depth independently defines the inconsistent subset; distributions are normalized separately.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 8: Target-view angle distributions. Dashed lines mark the mean minimum angle to input views:

12.34^{\circ}

for local sampling and

17.97^{\circ}

for ours.

Method	PSNR $\uparrow$ (dB)	SSIM $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$
AnySplat	15.6668	0.5531	0.4353	0.1939
C3G	13.2165	0.5653	0.6454	0.3267
QuerySplat	17.0532	0.6404	0.4571	0.2155
DA3-GS	17.7144	0.6463	0.4154	0.2077
WorldMirror 2.0	17.0031	0.6089	0.3757	0.1935
Ours	18.4002	0.6423	0.3026	0.1341

Appendix

Table 3: Pose-free results averaged over six datasets.

Method	PSNR $\uparrow$ (dB)	SSIM $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$
tttLRM	17.8632	0.6518	0.3793	0.2159
DA3-GS	16.8535	0.6100	0.4237	0.1965
DepthSplat	18.9085	0.6602	0.3718	0.1993
ReSplat	20.9139	0.7117	0.3333	0.1794
TokenGS	15.9149	0.5990	0.5779	0.2931
VolSplat	14.2476	0.5060	0.5192	0.2756

Appendix

Table 4: Pose-conditioned results averaged over six datasets.

Input	Method	DL3DV		ETH3D		NRGBD		Mip-NeRF 360		RE10K		DTU		Mean
		LPIPS $\downarrow$	DISTS $\downarrow$	LPIPS $\downarrow$	DISTS $\downarrow$	LPIPS $\downarrow$	DISTS $\downarrow$	LPIPS $\downarrow$	DISTS $\downarrow$	LPIPS $\downarrow$	DISTS $\downarrow$	LPIPS $\downarrow$	DISTS $\downarrow$	LPIPS $\downarrow$	DISTS $\downarrow$
Pose-free	AnySplat	0.6004	0.2527	0.5244	0.2245	0.1917	0.1153	0.5259	0.2187	0.2386	0.1084	0.4995	0.2452	0.4301	0.1941
	C3G	0.7565	0.3785	0.6090	0.2877	0.6137	0.3191	0.6849	0.3448	0.5080	0.2671	0.7031	0.3680	0.6459	0.3275
	QuerySplat	0.5453	0.2191	0.5315	0.2425	0.2580	0.1410	0.5621	0.2429	0.2172	0.1117	0.5243	0.2759	0.4397	0.2055
	DA3-GS	0.5347	0.2145	0.4937	0.2291	0.3352	0.1703	0.4933	0.2199	0.2412	0.1393	0.4287	0.2411	0.4211	0.2024
	WorldMirror 2.0	0.4647	0.2021	0.4484	0.2107	0.1793	0.1386	0.4432	0.1957	0.1903	0.1166	0.3924	0.2076	0.3530	0.1786

Appendix

Table 5: Per-dataset results with

N=4

input views. Mean weights datasets equally.

Panel of Figure 9 — Figure 9: Positional offsets on NRGBD with eight input views. Top: RGB; bottom: depth. Rendered depth is alpha-normalized expected camera- $z$ depth, without fitting to target GT. Colors indicate near (purple), far (yellow), and invalid (gray) pixels.

Variant	PSNR $\uparrow$ (dB)	SSIM $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$
Ours	16.9643	0.5836	0.4339	0.1661
Ours + Adam	16.9890	0.5864	0.4312	0.1656
Ours ∗	17.6074	0.6067	0.4176	0.1631

Appendix

Table 8: Three-step refinement on ETH3D, averaged over

N=4,8,16

input views.

Input resolution ( $W\times H$ )	Mean (s)	Min–max (s)
1022 $\times$ 756	5.619	5.617–5.621
756 $\times$ 560	2.533	2.532–2.535
504 $\times$ 378	1.032	1.031–1.034
378 $\times$ 280	0.564	0.563–0.564

Appendix

Table 9: Feed-forward network latency: mean and min–max over five runs.

Panel of Figure 11 — Figure 11: Pose-free rendering on ETH3D: grass textures and relief details.

Figure 12: Feed-forward reconstructions and renderings with 12–64 input views.

Figure 13: Large-scale reconstruction on KITTI. Insets show elevated-view renderings at locations A–F.

3D Gaussian Splatting

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

See all 986 papers

InfiniSplat: Implicit Gaussian Decoding for Large-Baseline Monocular View Synthesis
Aug 3, 2026Jiawei Wang, Hao Yu, Yongzhen Hu +73D Gaussian Splatting Novel View Synthesis
State Key Laboratory of CAD&CG, Zhejiang University, China · Zhejiang University, China · Udeer.ai, China +1
ViewSplat: View-Adaptive 3D Gaussian Splatting for Feed-Forward Synthesis
Mar 26, 2026Moonyeon Jeong, Seunggi Min, Suhyeon Lee +13D Gaussian Splatting Novel View Synthesis
University of Seoul · Korea Electronics Technology Institute
GIFSplat: Generative Prior-Guided Iterative Feed-Forward 3D Gaussian Splatting from Sparse Views
Feb 26, 2026Tianyu Chen, Wei Xiang, Kang Han +43D Reconstruction Sparse-View 3D Reconstruction
La Trobe University, Melbourne, VIC 3086, Australia · Cisco Research, San Jose, CA, USA

Aug 3, 2026J/K move · Enter open · S save

InfiniSplat: Implicit Gaussian Decoding for Large-Baseline Monocular View Synthesis

Jiawei Wang, Hao Yu, Yongzhen Hu, Xinyi Yang+6

State Key Laboratory of CAD&CG, Zhejiang University, China · Zhejiang University, China · Udeer.ai, China+1

3D Gaussian Splatting Novel View Synthesis Single-View 3D Reconstruction Feed-Forward 3D Reconstruction

Open paper AskPDF arXiv

Single-image feed-forward 3D Gaussian Splatting (3DGS) aims to directly generate a renderable 3D scene representation from one input image, avoiding the cost of multi-view capture and per-scene optimization. However, existing methods are often constrained by a pixel-aligned representation, where Gaussians are predicted from fixed image-grid locations. Such pixel-aligned primitives can produce promising nearby-view renderings, but they remain weakly coupled to underlying scene surfaces and struggle to preserve coherent structures under large viewpoint shifts. We present InfiniSplat, a feed-forward single-image 3DGS framework that moves from a pixel-aligned representation toward a surface-aligned representation. InfiniSplat constructs this representation by first using geometry-guided sampling to place 2D supports according to depth-induced local surface structure, and then applying a query-conditioned implicit decoder to predict Gaussian attributes from the image features queried at these supports. By grounding support locations in geometry while decoupling Gaussian prediction from fixed pixel centers, InfiniSplat produces Gaussian layouts that better follow scene surfaces and reduce scattered primitives caused by grid discretization. Across multiple cross-dataset NVS evaluations, InfiniSplat achieves state-of-the-art performance compared with single-image feed-forward baselines, and demonstrates zero-shot generalization from Hypersim indoor synthetic training to complex open-world scenes. Project page: https://zju3dv.github.io/InfiniSplat.

Abstract

3D Gaussian Splatting

Explore similar work

InfiniSplat: Implicit Gaussian Decoding for Large-Baseline Monocular View Synthesis

ViewSplat: View-Adaptive 3D Gaussian Splatting for Feed-Forward Synthesis

GIFSplat: Generative Prior-Guided Iterative Feed-Forward 3D Gaussian Splatting from Sparse Views