Organizations: State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, Beijing, China · School of Computer Science and Engineering, Beihang University, Beijing, China · Zhongguancun Laboratory, Beijing, China
3D Gaussian Splatting has achieved remarkable success in photorealistic rendering, yet it suffers from severe overfitting and geometric artifacts in sparse-view scenarios due to the inherent deficiency of photometric supervision. Recent advances have attempted to regularize optimization by incorporating external priors, such as depth, point clouds, or diffusion models. However, these methods typically overlook the non-uniform distribution of supervision across the viewing space, resulting in limited specificity in prior use and primitive control. In this paper, we propose GSPotential, a framework that quantifies view-space supervision imbalance using a Camera Potential Field. Our key insight is to identify supervision valleys where photometric constraints are most deficient, and use the potential field to guide reconstruction from two complementary aspects. First, we devise a probabilistic spherical sampling strategy that places informative virtual cameras in low-potential regions. Point-cloud renderings from these views then provide targeted geometric guidance. Second, the same field provides a directional coverage cue for conservative Gaussian updates in weakly covered spatial sectors. Extensive experiments demonstrate that GSPotential achieves high reconstruction fidelity while maintaining competitive training efficiency.
Figures & tables
Figure 1: Visualizations of scene reconstruction (inputs are from GaussianObject [ 30 ] ). Given sparse-view inputs, our GSPotential achieves improved reconstruction quality compared to original 3DGS. The boxes highlight regions with visible improvements.
Figure 2: Overview of GSPotential. Given sparse images, camera poses, and point clouds, we first estimate a Camera Potential Field to model view-space supervision imbalance. Based on this field, we perform Potential-Guided Viewpoint Generation to generate virtual cameras in low-potential regions. The sampled views are then used for Potential-Guided Reconstruction , while the projected potential provides a directional coverage score for Coverage-Aware Primitive Evolution, including anisotropy pruning and conservative densification.
Figure 3: Camera potential fields driven by sampling. We visualize the PSNR of 3DGS rendered images from random viewpoints across view space. (Left) In a single-camera setup, high-quality reconstruction is limited to a narrow cone. (Right) With 4 sparse cameras, despite the increase in coverage, significant supervision valleys (low PSNR zones) persist across viewpoints.
Figure 4: Comparison of kernel choices for modeling camera-potential decay. The horizontal axis θ denotes the angular distance between a novel viewpoint and its nearest training view. We compare vMF with κ=3.0 (left), sech with σ=5.3 (middle), and vMF with κ=4.4 (right). The latter two are both calibrated to decay to 1% of their peak at θ=90∘ , while κ=3.0 is selected to approximate the overall decay trend of the sech kernel. Compared with vMF, the sech kernel yields a smoother relation between camera potential and novel-view rendering quality.
Figure 5: Analysis of camera-potential parameterization and novel-view rendering quality. Each scatter point denotes a sampled viewpoint. The horizontal axis is the camera potential computed with a specific parameter setting, and the vertical axis reports the corresponding novel-view rendering quality. Darker colors indicate larger angular distance to the nearest training view. Compared with using angular distance alone, the potential value provides a more discriminative estimate of supervision strength, since viewpoints with similar nearest-view distance can still exhibit different rendering quality under different camera configurations. This verifies that the camera potential captures the joint camera-layout effect beyond a nearest-angle cue.
Dataset
Mip-NeRF 360
OmniObject3D
Method ∣ Metric
PSNR ↑
SSIM ↑
LPIPS ↓
TIME/s ↓
PSNR ↑
SSIM ↑
LPIPS ↓
TIME/s ↓
3DGS [ 12 ]
13.52
0.282
0.588
500
20.87
0.805
0.221
314
SparseGS [ 28 ]
11.15
0.166
0.672
627
19.15
0.751
0.305
745
FSGS [ 33 ]
13.49
0.324
0.573
762
20.13
0.797
0.206
575
CoR-GS [ 31 ]
12.90
0.309
0.574
2110
21.44
0.813
0.197
1559
SplatFields [ 16 ]
11.92
0.331
0.748
2865
21.12
0.811
0.235
1194
Table 1: Quantitative evaluation of our method compared to previous works on two datasets. Results on Mip-NeRF 360 are averaged over seven scenes, and all metrics are computed on uniformly sampled test views. (“–” denotes unavailable results.)
Figure 6: Comparison of 3DGS training with different point cloud initialization strategies. The plots visualize training iterations (x) versus PSNR (y), with: left for original point cloud (generated by DUSt3R), 1/4, 1/16, and 1/64 down-sampling ratios; right for original point cloud (generated by DUSt3R) and random 3D noise initialization with the same down-sampling ratios.
Figure 7: Qualitative comparison on unsupervised views. Boxes highlight regions with notable improvements by GSPotential.
Lvirt
Uinfo
Qi -P
Qi -D
Uconf
PSNR ↑
SSIM ↑
LPIPS ↓
×
×
×
×
×
13.52
0.282
0.588
✓
×
×
×
×
14.49
0.356
0.624
✓
✓
×
×
×
15.90
0.379
0.601
✓
✓
✓
×
×
16.03
0.388
0.585
✓
✓
×
✓
×
16.17
0.384
0.594
✓
✓
✓
✓
×
16.34
0.394
0.577
Table 2: Ablations of different components of GSPotential, evaluated on Mip-NeRF 360. Uinfo denotes the information potential used for low-potential viewpoint generation. Qi -P and Qi -D denote coverage-aware pruning and densification guided by the directional coverage score Qi , respectively. Uconf denotes the confidence term for suppressing unreliable virtual supervision near extrapolation regions.
Figure 8: Qualitative ablation on test view of the Mip-NeRF 360 counter scene. Columns correspond to the variants in Table 2 .
Views
Ours
3DGS
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
1
15.73
0.392
0.571
13.02
0.281
0.604
2
16.14
0.408
0.527
13.21
0.282
0.597
4
16.43
0.411
0.505
13.52
0.282
0.588
8
16.59
0.415
0.480
13.78
0.282
0.581
Table 3: Ablation of point cloud quality, evaluated on Mip-NeRF 360. Views refer to number of views used to build the point cloud.
Sampling
PSNR ↑
SSIM ↑
LPIPS ↓
Nearest-angle random
15.63
0.318
0.520
Uniform sphere
15.58
0.306
0.525
Farthest-point sphere
15.50
0.305
0.539
Uniform interpolation
15.52
0.310
0.522
CPF (ours)
15.93
0.373
0.468
Table 4: Controlled comparison of virtual-view sampling strategies on the Mip-NeRF 360 kitchen scene. All variants use the same reconstruction setting and differ only in virtual-view selection.
Views
3DGS
GSPotential
3
14.81/0.331/0.508
15.49/0.358/0.478
4
15.09/0.337/0.495
15.93/0.373/0.468
5
15.78/0.398/0.460
17.00/0.412/0.441
6
15.59/0.359/0.495
15.89/0.413/0.464
9
16.81/0.469/0.441
17.19/0.498/0.404
Table 5: Sensitivity to the number of input views on the Mip-NeRF 360 kitchen scene. Each entry reports PSNR/SSIM/LPIPS.
3D Gaussian Splatting (3DGS) has recently enabled real-time rendering of unbounded 3D scenes for novel view synthesis. However, this technique requires dense training views to accurately reconstruct 3D geometry. A limited number of input views will significantly degrade reconstruction quality, resulting in artifacts such as "floaters" and "background collapse" at unseen viewpoints. In this work, we introduce SparseGS, an efficient training pipeline designed to address the limitations of 3DGS in scenarios with sparse training views. SparseGS incorporates depth priors, novel depth rendering techniques, and a pruning heuristic to mitigate floater artifacts, alongside an Unseen Viewpoint Regularization module to alleviate background collapses. Our extensive evaluations on the Mip-NeRF360, LLFF, and DTU datasets demonstrate that SparseGS achieves high-quality reconstruction in both unbounded and forward-facing scenarios, with as few as 12 and 3 input images, respectively, while maintaining fast training and real-time rendering capabilities.
Generating high-quality novel views at real-time frame rates remains a central challenge in 3D vision, particularly in sparse-view scenarios. Neural radiance fields have demonstrated robust reconstruction from limited observations, but their reliance on volumetric rendering leads to high computational cost and slow inference. In contrast, Gaussian Splatting methods achieve real-time rendering through rasterization, but their optimization is highly sensitive to the quality of the initial geometry. This sensitivity becomes especially problematic in sparse-view settings, where limited observations often lead to incomplete or noisy point-cloud reconstructions. In this work, we present AugSplat, a simple framework for improving Gaussian Splatting in sparse-view regimes using radiance-field-based view augmentation. We first train a radiance field on the sparse input views and use it to synthesize additional images from nearby novel viewpoints, increasing the effective view-space coverage available for supervision. These synthetic views are then used as auxiliary supervision during Gaussian Splatting optimization. We study two variants: Staged AugSplat, which uses synthetic views for an initial optimization phase before switching to real images, and Dual AugSplat, which jointly trains on real and synthetic views with a decaying synthetic loss weight. Experiments on sparse-view mip-NeRF 360 scenes show that AugSplat improves reconstruction quality over standard Gaussian Splatting. Staged AugSplat achieves the strongest average performance, while Dual AugSplat provides a closely performing formulation that keeps real-image supervision active throughout training, and both variants preserve real-time rendering at inference.
Lorenzo Lazzaroni, Riccardo Bollati, Daniel Barath +2
Recent advances in neural rendering have established 3D Gaussian Splatting (3DGS) as a highly efficient representation for novel view synthesis, enabling fast training and real-time rendering with strong fidelity. However, when supervision is limited to sparse input views, 3DGS tends to overfit to the observed images and generalize poorly to unseen viewpoints. We address this challenge from the perspective of flat minima (FM) optimization, which seeks solutions that remain stable under small parameter perturbations. Viewing Gaussian parameters as trainable weights, we adapt FM principles to the geometric and dynamic nature of 3DGS with a lightweight training framework. Our method regularizes optimization with controlled Gaussian perturbations that account for each Gaussian's anisotropy and the training progress, preserving fine details while improving robustness to sparse-view overfitting. To further stabilize this flat minima optimization process, we introduce periodic reinitialization, which temporarily returns non-positional parameters to their initial states for a short window. Together, these techniques integrate seamlessly into existing 3DGS pipelines without architectural changes. Experiments on LLFF and Mip-NeRF360 datasets demonstrate improved quantitative metrics and perceptual quality under sparse-view supervision, producing reconstructions that are sharper, more stable, and better generalized to novel viewpoints.