The emergence of 3D Gaussian Splatting has fundamentally redefined the capabilities of photorealistic neural rendering by enabling high-throughput synthesis of complex environments. While procedural methods like Wang Tiles have recently been integrated to facilitate the generation of expansive landscapes, these systems typically remain constrained by a reliance on densely sampled exemplar reconstructions. We present DAV-GSWT, a data-efficient framework that leverages diffusion priors and active view sampling to synthesize high-fidelity Gaussian Splatting Wang Tiles from minimal input observations. By integrating a hierarchical uncertainty quantification mechanism with generative diffusion models, our approach autonomously identifies the most informative viewpoints while hallucinating missing structural details to ensure seamless tile transitions. Experimental results indicate that our system significantly reduces the required data volume while maintaining the visual integrity and interactive performance necessary for large-scale virtual environments.
Figures & tables
Figure 1: Overview of the DAV-GSWT framework for data-efficient Gaussian Splatting and tiling. The pipeline begins with a coarse reconstruction G0 computed from sparse initial images Iinit . During the active cycle, a pre-trained diffusion model generates M stochastic latent samples zm(θ) using attention dropout. These samples are evaluated by the uncertainty estimator, which computes a score u(θ) from image-space LPIPS gradients or the latent 2-Wasserstein divergence W2(Z) . The top- k poses Θ∗ are selected for physical acquisition to refine the field into GT . In the synthesis stage, the refined field is partitioned into Wang tiles T , and seam continuity is optimized through an uncertainty-adaptive graph cut that adjusts the semantic weight γ(uˉ) to maintain perceptual and geometric consistency.
Dataset
Splat count (M)
Exemplar recon. time (min)
Tile Constr. time (s)
Pre-sorting time (s)
Render time (ms)
Sort time (ms)
Update time (ms)
Desert (Synth.)
5.5
68.5
76.6
3.54
5.17 ± 0.54
4.70 ± 1.23 (92.15%)
4.15 ± 0.30 (9.83%)
Flowers (Synth.)
17.9
66.9
97.5
5.39
12.26 ± 3.20
9.46 ± 3.51 (97.25%)
3.53 ± 0.30 (22.74%)
Grass (Synth.)
16.2
67.6
86.3
5.86
11.35 ± 2.56
12.13 ± 4.66 (83.12%)
3.63 ± 0.34 (20.18%)
Planet (Synth.)
11.7
65.0
100.6
8.97
8.88 ± 2.01
7.56 ± 3.33 (91.45%)
3.51 ± 0.42 (16.82%)
Meadow (Synth.)
22.5
65.0
102.0
8.43
14.90 ± 3.61
15.46 ± 6.05 (83.41%)
3.68 ± 0.29 (26.91%)
Forest (Real)
6.4
45.9
71.8
3.03
5.67 ± 0.60
4.18 ± 1.22 (98.02%)
3.68 ± 0.29 (10.95%)
Table 1: Real-time tiling and rendering statistics for DAV-GSWT across different scenes. All scenes are parameterized on a random height field with constant camera movement at fixed height.
Dataset
LOD 0
LOD 1
LOD 2
LOD 3
LOD 4
LOD 5
Desert
0.00319
0.0084
0.0223
0.0501
0.1029
0.2023
(Synth.)
64.2K
17.3K
4.42K
1.12K
279
60.6
Flowers
0.0077
0.0142
0.0251
0.0388
0.0698
0.1337
(Synth.)
83.3K
19.7K
4.73K
1.37K
344
84.7
Grass
0.0085
0.0150
0.0278
0.0454
0.0812
0.1568
(Synth.)
87.5K
20.9K
4.97K
1.13K
270
70.7
Table 2: Average Gaussian scale factor (top) and Gaussian count per tile (bottom) at different LODs for DAV-GSWT datasets.
Figure 2: Active-view uncertainty over a dense candidate viewing sphere. Each point represents a candidate camera pose parameterized by azimuth and elevation, colored by diffusion epistemic uncertainty. White star markers indicate the top- k views selected for physical capture.
Figure 3: Iterative reconstruction evolution under DAV-GSWT. Top row shows rendered reconstructions at iterations T=0,1,3 . Bottom row visualizes the corresponding absolute error maps with respect to the final reconstruction. Error magnitudes are normalized and share a common color scale.
Figure 4: Ablation study of uncertainty formulations for active view selection. From left to right and top to bottom: image-gradient-based uncertainty with LPIPS, Wasserstein-2 only, Wasserstein-2 combined with LPIPS, and ground truth. Seam-level LPIPS scores are reported for each variant.
Figure 5: Comparison of seam artifacts using color-only graph cuts versus semantic-aware cuts augmented with SAM. Semantic constraints significantly reduce seam density around object boundaries.
Figure 6: Visualization of the tile-level uncertainty cache during online reconstruction. Warm colors indicate high-uncertainty regions prioritized for refinement, while cool colors denote confident tiles. The overlaid path illustrates the adaptive camera traversal.
Figure 7: Reconstruction quality versus capture budget. DAV-GSWT achieves near-exhaustive reconstruction quality with substantially fewer captured views compared to random and exhaustive strategies.
Parameter
Value
PSNR ↑
Seam-LPIPS ↓
Unc. Time ↓ (s)
k
10
28.91 ± 0.09
0.036 ± 0.003
0.9 ± 0.1
20
29.41 ± 0.08
0.031 ± 0.002
1.8 ± 0.1
30
29.45 ± 0.08
0.030 ± 0.002
2.7 ± 0.2
pdrop
0.05
28.76 ± 0.10
0.038 ± 0.003
1.8 ± 0.1
0.15
29.41 ± 0.08
0.031 ± 0.002
1.8 ± 0.1
0.30
29.22 ± 0.09
0.033 ± 0.003
1.8 ± 0.1
Table 3: Sensitivity of DAV-GSWT to core hyper-parameters. Defaults in bold.
Configuration
PSNR ↑
Seam-LPIPS ↓
VRAM ↓ (GB)
Render ↓ (ms)
Unc-Time ↓ (s)
p -value
Full ensemble W 2 +LPIPS
29.41 ± 0.08
0.031 ± 0.002
6.2 ± 0.1
7.4 ± 0.3
1.8 ± 0.1
—
W 2 only
29.12 ± 0.09
0.034 ± 0.003
5.9 ± 0.1
7.2 ± 0.2
1.3 ± 0.1
0.006
Image-space grad+LPIPS
28.55 ± 0.11
0.039 ± 0.003
8.1 ± 0.2
9.1 ± 0.4
4.7 ± 0.2
<0.001
Single forward (no dropout)
27.03 ± 0.15
0.051 ± 0.004
5.6 ± 0.1
6.9 ± 0.2
0.3 ± 0.0
<0.001
Graph-cut w/o γ(uˉ)
28.61 ± 0.10
0.043 ± 0.003
6.0 ± 0.1
7.5 ± 0.3
—
0.002
Exhaustive 200-view
29.50 ± 0.07
0.030 ± 0.002
11.3 ± 0.2
12.6 ± 0.5
—
0.18
Table 4: Ablation of uncertainty estimator and seam-energy variants. Top- k =20 views/iter, 5 runs. ↑ higher is better, ↓ lower. p<0.01 vs. next best in each block (paired t -test). Seam-LPIPS measured on 512 × 512 boundary crops. Two-tailed paired t-test, t=4.28 , df=4 , d=1.92 .
3D Gaussian Splatting (3DGS) achieves high-quality novel-view synthesis by optimizing freely placed primitives in 3D and adaptively densifying them in under-reconstructed regions. However, this scene-adaptive capacity allocation is largely lost in existing feed-forward 3DGS methods, which commonly regress Gaussians at input pixels and lift them along camera rays. Such pixel-aligned formulations make the number and placement of primitives depend on image resolution and input viewpoints rather than scene complexity, resulting in dense and often redundant Gaussian sets. We present ATSplat, a feed-forward 3DGS framework that restores the adaptive allocation capability of 3DGS optimization through Adaptive 3D Tokens. ATSplat first lifts coarse patch-level depth and camera cues into sparse 3D anchor tokens, forming a compact scaffold of the scene. Each token is then regressed into local Gaussians with learnable 3D offsets, decoupling primitive placement from input image grids. An Adaptive Token Expansion module predicts a token-level uncertainty score, supervised by rendering error maps, and selectively expands high-uncertainty tokens through learnable expansion layers. This sparse-to-adaptive formulation enables ATSplat to concentrate primitives in challenging regions while maintaining a compact representation. Experiments on two representative datasets, RealEstate10K and DL3DV, show that ATSplat achieves state-of-the-art rendering quality while reducing the number of Gaussians by more than 5.7× compared with dense feed-forward 3DGS methods. From 12 input images at 512×960 resolution, ATSplat completes reconstruction in less than a second using a single commercial GPU, and renders high-quality novel views at 1136 FPS (512×960) with only 311K Gaussians.
In Cho, Jeonghwan Cho, Mijin Yoo +2
Yonsei University, South Korea · National University of Singapore, Singapore
3D Gaussian Splatting (3DGS) has emerged as an advanced technique for real-time novel view synthesis by representing scene geometry and appearance using differentiable Gaussian primitives. However, efficiently computing precise Gaussian-tile intersections remains a critical task in the rasterization pipeline. To this end, we propose QuadBox, a method that leverages four axis-aligned bounding boxes to tightly encapsulate projected Gaussians in a discrete manner. First, we derive a geometry-aware stretching factor that enables the construction of a tile-aligned QuadBox, which covers the elliptical projection and largely excludes irrelevant tiles. Second, we introduce QPass, a single-pass tile traversal algorithm that exhaustively exploits the discrete nature of QuadBox, ensuring that the tile intersection check is performed with simple interval tests. Experiments on public datasets show that our method accelerates the rendering speed of 3DGS by 1.85×. Code is available at https://github.com/Powertony102/QuadBox.
Xinze Li, Bohan Yang, Pengxu Chen +4
Beijing Normal-Hong Kong Baptist University · Jilin University · Hong Kong Baptist University +2
3D Gaussian Splatting has revolutionized neural rendering with real-time performance. However, scaling this approach to large scenes using Level-of-Detail methods faces critical challenges: inefficient serial traversal consuming over 60% of rendering time, and redundant Gaussian-tile pairs that incur unnecessary processing overhead. To address these limitations, we introduce FilterGS, featuring a parallel filtering mechanism with two complementary filters that select Gaussian elements efficiently without tree traversal. Additionally, we propose a novel GTC metric that quantifies the redundancy of Gaussian-tile key-value pairs. Based on this metric, we introduce a scene-adaptive Gaussian shrinking strategy that effectively reduces redundant pairs. Extensive experiments demonstrate that FilterGS achieves state-of-the-art rendering speeds while maintaining competitive visual quality across multiple large-scale datasets. Project page: https://github.com/xenon-w/FilterGS