CylinderSplat: 3D Gaussian Splatting with Cylindrical Triplanes for Panoramic Novel View Synthesis

Qiwei Wang; Xianghui Ze; Jingyi Yu; Yujiao Shi

cs.CVMar 6, 2026

CylinderSplat: 3D Gaussian Splatting with Cylindrical Triplanes for Panoramic Novel View Synthesis

Authors: Qiwei Wang, Xianghui Ze, Jingyi Yu, Yujiao Shi

Organizations: Shanghaitech University · Nanjing University of Science and Technology

Abstract

Feed-forward 3D Gaussian Splatting (3DGS) has shown great promise for real-time novel view synthesis, but its application to panoramic imagery remains challenging. Existing methods often rely on multi-view cost volumes for geometric refinement, which struggle to resolve occlusions in sparse-view scenarios. Furthermore, standard volumetric representations like Cartesian Triplanes are poor in capturing the inherent geometry of $360^\circ$ scenes, leading to distortion and aliasing. In this work, we introduce CylinderSplat, a feed-forward framework for panoramic 3DGS that addresses these limitations. The core of our method is a new {cylindrical Triplane} representation, which is better aligned with panoramic data and real-world structures adhering to the Manhattan-world assumption. We use a dual-branch architecture: a pixel-based branch reconstructs well-observed regions, while a volume-based branch leverages the cylindrical Triplane to complete occluded or sparsely-viewed areas. Our framework is designed to flexibly handle a variable number of input views, from single to multiple panoramas. Extensive experiments demonstrate that CylinderSplat achieves state-of-the-art results in both single-view and multi-view panoramic novel view synthesis, outperforming previous methods in both reconstruction quality and geometric accuracy.

Figures & tables

Figure 1: This paper introduces CylinderSplat, a feed-forward panoramic 3D Gaussian Splatting (3DGS) framework for panoramic novel view synthesis from single (left) or sparse (right) input views.

Figure 2: Visualization of the Triplane representation in (a) Cartesian, (b) Spherical, and (c) Cylindrical coordinate systems. (d) The corresponding unit volume elements for each system.

Figure 3: Overview of our CylinderSplat framework . Our method uses a dual-branch architecture trained via a three-stage curriculum. The pixel branch uses a multi-view attention mechanism to generate high-quality Gaussians for well-observed regions. The volume branch is designed to fill the gaps by lifting features into our cylindrical triplane representation, thereby completing the scene geometry robustly. The outputs from both branches are then unified for a final render.

Panel of Table 1 — Table 1: Quantitative comparison for the two-view reconstruction on the Matterport3D, Replica, and Residential datasets. The first , second , and third best results are highlighted. Methods marked with * indicate that we reimplemented them using their official code.

Panel of Figure 5 — Figure 5: Qualitative comparison of synthetic scenes for the two-view input task across different baselines. The top two rows display the target ground truth and input views, followed by comparisons of different methods. Zoomed-in regions highlight the superior completeness, sharpness, and reduced artifacts of our method.

Table 6

Figure 7: Ablation study visualizations on Matterport3D (leftmost column is ground truth). Among the different coordinate systems for the Triplane, the Cartesian version shows significant distortion, and the spherical version struggles with distant rooms. Our cylindrical Triplane performs best. While the pixel branch is high-quality in visible areas, combining it with the Triplane branch improves reconstruction of distant regions.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Configuration $(r,z,\theta)$	PCC $\uparrow$	WS-PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Memory	Time
$(N_{r},N_{z},N_{\theta})=(16,32,64)$	0.72	21.18	0.71	0.24	15GB	0.26s
$(N_{r},N_{z},N_{\theta})=(16,64,128)$	0.78	22.17	0.78	0.21	17GB	0.29s
$(N_{r},N_{z},N_{\theta})=(16,128,256)$	0.80	23.88	0.82	0.19	25GB	0.34s
$(N^{\prime}_{r},N^{\prime}_{z},N^{\prime}_{\theta})=(8,16,32)$	0.77	22.08	0.73	0.28	16GB	0.27s
$(N^{\prime}_{r},N^{\prime}_{z},N^{\prime}_{\theta})=(8,32,64)$	0.78	22.17	0.78	0.21	17GB	0.29s
$(N^{\prime}_{r},N^{\prime}_{z},N^{\prime}_{\theta})=(8,64,128)$	0.79	22.46	0.80	0.21	18GB	0.31s

Appendix

Table 8: Ablation study on sampling hyperparameters (

(N_{r},N_{z},N_{\theta})

and

(N^{\prime}_{r},N^{\prime}_{z},N^{\prime}_{\theta})

) on the Matterport3D (2.0m baseline) dataset with only the volume branch.

Figure 9: Visualization of our Triplane initialization strategies. In this figure, a blue triangle represents Camera A, with blue circles denoting its corresponding feature point cloud. A yellow triangle represents Camera B with its yellow point cloud. We construct a local cylindrical Triplane at each camera’s location and initialize it by projecting feature points onto its three planes. (a) For static scenes, the Triplane for Camera A is formed by aggregating all candidate feature points that fall within its volume, regardless of their origin (both blue and yellow points). (b) In contrast, for scenes with dynamic objects, the Triplane for Camera A is formed exclusively by feature points from its own view (only blue points). This strategy prevents dynamic inconsistencies from one view (e.g., yellow points from Camera B) from corrupting another view’s Triplane with noise.

Dataset	Matterport3D (2.0m baseline)				360Loc (avg. 1.40m baseline)
Method	PCC $\uparrow$	WS-PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PCC $\uparrow$	WS-PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Shared Initialization	0.851	23.76	0.835	0.175	0.831	27.33	0.841	0.107
Isolated Initialization	0.847	23.24	0.796	0.197	0.883	28.35	0.896	0.095

Appendix

Table 9: Comparison of different Triplane initialization strategies on static (Matterport3D) and dynamic (360Loc) datasets.

Dataset	Matterport3D
Baseline	2.0m				1.5m				1.0m				Inference Time
Render Method	PCC $\uparrow$	WS-PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PCC $\uparrow$	WS-PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PCC $\uparrow$	WS-PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
CubeMap	0.815	23.71	0.822	0.171	0.853	25.51	0.856	0.129	0.906	28.86	0.936	0.076	0.33s
Panorama Gaussian	0.851	23.76	0.835	0.175	0.867	25.91	0.873	0.128	0.923	28.89	0.937	0.081	0.29s

Appendix

Table 10: Comparison of different rendering methods on the Matterport3D dataset for the two-view input task, evaluating quality metrics and inference time.

Figure 10: Visualizing the advantages of the cylindrical Triplane. (a) The shape of the sampling volumes for each coordinate system when fitting a typical synthetic scene, highlighting their geometric alignment. (b) Feature distribution during Triplane initialization. At this stage, panoramic feature point clouds are projected onto the Triplane surfaces. The Cartesian system suffers from heavy point overlap, where many distinct 3D feature points are projected to the same grid cells. In contrast, the spherical and cylindrical systems achieve a much more even distribution of features. (c) Projection patterns for Triplane-to-Image Attention. We visualize how sample points from each Triplane’s volume project onto the panorama for cross-attention. This is shown across six maps for each coordinate system, ordered by increasing distance from the origin. For the Cartesian system (left), maps show that on a Cartesian plane, as this plane moves farther from the origin, its panoramic projection exhibits a limited field of view and severe distortion. For the spherical system (middle), maps illustrate that on a spherical surface, with increasing sphere radius (distance from origin), its projection remains consistently uniform across the panorama. Finally, for the cylindrical system (right), maps display that on a cylindrical surface, as its radius increases, the projection remains relatively uniform but naturally focuses more on the panoramic center, better aligning with real-world depth distributions. (d) Qualitative rendering comparison using only the volume branch, demonstrating the superior performance of the cylindrical representation in novel view synthesis.

Method	AbsRel $\downarrow$	RMSE $\downarrow$	$\delta_{1}$ $\uparrow$	PCC $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
OmniScene	0.36	0.77	0.37	0.73	22.75	0.71	0.24
Splatter360	0.47	0.91	0.28	0.68	21.31	0.74	0.28
PanSplat	0.41	1.08	0.11	0.71	20.56	0.77	0.26
Ours (DepthAnywhere)	0.34	0.66	0.60	0.78	23.08	0.78	0.22
Ours (UniFuse)	0.31	0.64	0.62	0.78	23.31	0.81	0.21
Ours (UniK3D)	0.20	0.48	0.75	0.85	23.76	0.83	0.17

Appendix

Table 11: Different depth prior on Matterport3D 2.0m baseline.

Method	abs_rel $\downarrow$	rmse $\downarrow$	delta1 $\uparrow$	PCC $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Ours (Cartesian Scale)	0.22	0.55	0.73	0.82	23.54	0.82	0.18
Ours (Cylinder Scale)	0.20	0.48	0.75	0.85	23.76	0.84	0.18

Appendix

Table 12: Ablation study on scale transformation strategy (Matterport3D, 2.0m baseline).

	Kansas Dataset (20m–30m)				360Loc (4.5m Baseline)				360Loc (3.0m Baseline)
Method	PCC $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PCC $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PCC $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
OmniScene	0.55	17.08	0.44	0.35	0.84	18.12	0.53	0.41	0.86	21.83	0.66	0.30
Splatter360	0.39	16.21	0.36	0.47	0.80	19.93	0.62	0.32	0.82	22.65	0.71	0.24
PanSplat	0.07	15.02	0.29	0.61	0.10	21.03	0.69	0.28	0.11	23.47	0.75	0.22
Ours	0.63	21.03	0.67	0.21	0.86	22.68	0.73	0.22	0.86	24.50	0.79	0.18

Appendix

Table 13: Quantitative comparison on wide-baseline real-world scenarios. We report results on the Kansas dataset (extreme 20m–30m baseline) and 360Loc dataset with increasing baselines (4.5m and 3.0m). Our method consistently outperforms baselines, with the performance gap widening as the difficulty increases.

Method	Training Time (per iteration)	Training Epochs
OmniScene	3.23s	10
Splatter360	2.89s	10
PanSplat	2.17s	10
Ours (Stage 1: Pixel Branch)	0.81s	10
Ours (Stage 2: Volume Branch)	1.13s	10
Ours (Stage 3: Joint Training)	1.98s	10

Appendix

Table 14: Comparison of training efficiency (time per iteration and epochs) against baselines.

Table 18

Method	AbsRel $\downarrow$	RMSE $\downarrow$	$\delta_{1}$ $\uparrow$	PCC $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Inference Time
OmniScene (Cubemap)	0.37	0.65	0.53	0.73	22.75	0.71	0.24	0.48s
OmniScene (Direct Raster)	0.29	0.63	0.61	0.76	22.19	0.76	0.24	0.44s

Appendix

Table 17: Comparison of OmniScene performance using Cubemap vs. Direct Panoramic Rasterizer on Matterport3D 2.0m baseline.

Figure 11: Qualitative comparisons on synthetic datasets for the two-view input task, with an input view baseline of 2.0m. The first column shows the rendered RGB image for a novel view, and the second column displays its corresponding depth map. The third and fourth columns provide zoomed-in results of the RGB image and depth map, respectively.

Figure 12: Qualitative comparisons on synthetic datasets for the single-view input task, where the distance between the input and output views is 1.0m. The first column shows the rendered RGB image for a novel view, and the second column displays its corresponding depth map. The third and fourth columns provide zoomed-in results of the RGB image and depth map, respectively.

Figure 13: Qualitative comparison on the real-world outdoor Kansas Dataset for the two-view input task. The distance between the two input views is approximately 20–30m, making it a highly challenging scenario. Although some blurriness remains, our method significantly outperforms the baselines, demonstrating the effectiveness of CylinderSplat for sparse-view synthesis under demanding real-world conditions.

Explore similar work

Sep 23, 2025

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

Feed-forward 3D Gaussian Splatting (3DGS) has emerged as a highly effective solution for novel view synthesis. Existing methods predominantly rely on a pixel-aligned Gaussian prediction paradigm, where each 2D pixel is mapped to a 3D Gaussian. We rethink this widely adopted formulation and identify several inherent limitations: it renders the reconstructed 3D models heavily dependent on the number of input views, leads to view-biased density distributions, and introduces alignment errors, particularly when source views contain occlusions or low texture. To address these challenges, we introduce VolSplat, a new multi-view feed-forward paradigm that replaces pixel alignment with voxel-aligned Gaussians. By directly predicting Gaussians from a predicted 3D voxel grid, it overcomes pixel alignment's reliance on error-prone 2D feature matching, ensuring robust multi-view consistency. Furthermore, it enables adaptive control over density based on 3D scene complexity, yielding more faithful Gaussians, improved geometric consistency, and enhanced novel-view rendering quality. Experiments on widely used benchmarks demonstrate that VolSplat achieves state-of-the-art performance, while producing more plausible and view-consistent results. The video results, code and trained models are available on our project page: https://lhmd.top/volsplat.

Weijie Wang, Yeqing Chen, Zeyu Zhang +9

Zhejiang University · GigaAI · The Chinese University of Hong Kong +4

Multi-View Consistency 3D Gaussian Splatting Novel View Synthesis Feed-Forward 3D Reconstruction

Ask PDF

May 13, 2026

PanoPlane: Plane-Aware Panoramic Completion for Sparse-View Indoor 3D Gaussian Splatting

We present PanoPlane, an approach for high-fidelity sparse-view indoor novel view synthesis that reconstructs closed room geometry via panoramic scene completion. Unlike perspective-based methods that generate training views from limited fields of view, PanoPlane leverages

360^{\circ}

panoramic completion to condition the generative process on the full spatial layout. We propose Layout Anchored Attention Steering, a training-free mechanism that steers attention within the diffusion model's internal representation toward scene's detected planar surfaces at inference time. By directing each unobserved region's attention toward geometrically consistent observed content, our method replaces unconstrained hallucination with grounded surface extrapolation. The resulting panoramic completions provide supervision for 3D Gaussian Splatting, enabling accurate novel-view synthesis across unobserved regions from as few as three input views. Experiments on Replica, ScanNet++, and Matterport3D demonstrate state-of-the-art novel view synthesis quality across 3, 6, and 9 input views, achieving up to

+17.8\%

improvement in PSNR over the current state-of-the-art baseline without any training or fine-tuning of the diffusion model.

Adil Qureshi, Dongki Jung, Jaehoon Choi +1

University of Maryland, College Park

3D Gaussian Splatting Diffusion Model Guidance Panoramic 3D Reconstruction Novel View Synthesis Sparse-View 3D Reconstruction Indoor Scene Reconstruction+1

Ask PDF

Jun 29, 2026

StereoGS: Sparse-View 3D Gaussian Splatting via Stereo Priors

3D Gaussian Splatting (3DGS) has achieved remarkable success in real-time novel view synthesis, yet it suffers from severe overfitting under sparse-view settings due to insufficient geometric constraints. While recent methods introduce monocular depth priors to mitigate this, they inherently struggle with scale ambiguity and cross-view inconsistency, leading to defective geometry. In this paper, we propose StereoGS, a novel sparse-view 3DGS framework that integrates stereo priors to establish reliable binocular consistency. Unlike scale-agnostic monocular constraints, StereoGS introduces a Stereo Depth Regularization by constructing virtual stereo pairs during optimization and leveraging a foundation stereo model to enforce absolute scale and binocular-consistent structures. To further suppress overfitting and eliminate redundant primitives, we design a Gradient-Aware Opacity Decay strategy that dynamically penalizes Gaussians based on their relative opacity gradient magnitudes. Combined with a Consistency-Aware Dense Initialization using zero-shot multi-view depth estimation, StereoGS effectively anchors primitives to accurate scene surfaces. Extensive experiments on LLFF, DTU, Mip-NeRF360, and Blender datasets demonstrate that StereoGS achieves state-of-the-art performance in sparse-view settings without incurring any additional inference overhead. Project Page: https://stringerywh00.github.io/StereoGS_project_page/

Wenhao Yuan, Yiyuan Ge, Deli Cai

South China University of Technology, China

Multi-View Consistency 3D Gaussian Splatting Sparse-View 3D Reconstruction

Ask PDF