We identify a key limitation of existing feed-forward 3DGS methods: Gaussian attributes are typically predicted in a radial-distance-normalized space that canonicalizes a global radial statistic of the reconstructed geometry, while the spacing between neighboring Gaussian centers can still vary substantially across scene scales and camera configurations. This mismatch induces systematic shifts in Gaussian-scale targets, leading to ambiguous supervision. To address this, we propose SplatAnything with Gaussian Spacing Normalization (GSNorm), which canonicalizes a robust scene-level statistic of Gaussian spacing to provide a more consistent spatial reference for Gaussian prediction. To rigorously assess our method, we further establish a systematic benchmark spanning diverse datasets, input sparsity levels, and challenging extrapolated viewpoints. Extensive experiments demonstrate that SplatAnything achieves state-of-the-art performance, with particularly strong gains under substantial variations in scene scale and viewpoint.
Figures & tables
Figure 1: Given any number of input images and optional camera poses, SplatAnything reconstructs a 3D Gaussian Splatting (3DGS) representation within seconds. It generalizes across arbitrary scene scales and camera configurations, while supporting challenging extrapolated viewpoints. Compared with prior feed-forward 3DGS methods, SplatAnything achieves substantially stronger cross-scene generalization and novel-view rendering quality.
Figure 2: Normalization space. (a) Radial Distance Normalization canonicalizes a global radial statistic but leaves Gaussian spacing varying across scenes and camera configurations, resulting in inconsistent supervision targets. (b) Gaussian Spacing Normalization canonicalizes a representative statistic of Gaussian spacing, providing a more consistent spatial reference for Gaussian prediction.
Figure 3: Architecture of SplatAnything. Given N input images and optional camera parameters, a frozen geometry foundation model first estimates multi-view geometry. We convert the geometry from its native radial-distance-normalized space to the proposed Gaussian-spacing-normalized space and predict per-pixel Gaussian attribute through a Gaussian backbone and a lightweight GShead.
Figure 4: Median Gaussian spacing across scene types. RDNorm leaves up to 4.5× variation in median Gaussian spacing across scene categories, whereas GSNorm canonicalizes it by construction.
Figure 5: Qualitative comparisons across diverse scenes and viewpoints with multi-view input.
Table 2: Quantitative comparison of feed-forward 3DGS methods using a single input view . The top-3 results are highlighted as first , second , and third .
Figure 7: Ablation studies on the effect of GSNorm across different scene types.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Data Type
Dataset
# Scenes
Resolution ( H×W )
Synthetic
HyperSim ( Roberts et al., 2021 )
364
768 × 1024
UnrealStereo4K ( Tosi et al., 2021 )
18
1080 × 1920
OmniObject3D ( Wu et al., 2023 )
5914
800 × 800
Rawmantic3D ( Team, 2026 )
30
1080 × 1920
OmniWorld ( Zhou et al., 2025 )
10464
720 × 1280
GTA-SfM ( Wang and Shen, 2020 )
200
480 × 640
Appendix
Table 4: Details of the training datasets used by SplatAnything.
Figure 8: Source-view trajectory patterns ( 16 views per panel), visualized in an approximately top-down projection. Dots and arrows denote camera centers and viewing directions, respectively; the gray polyline connects consecutive views, and color indicates capture order. For inward-orbit , × marks the approximate viewing center.
Figure 9: Target-view sampling modes used during training. Blue cameras denote the source views provided as input, while red cameras denote the target views used for rendering supervision.
Input
Method
4 Views
8 Views
16 Views
32 Views
64 Views
PSNR ↑
PSNR ↑
PSNR ↑
PSNR ↑
PSNR ↑
w/o Pose
AnySplat
12.68
13.20
12.83
13.12
14.26
YoNoSplat
18.23
18.64
18.73
18.34
15.79
DA3-GS
16.32
16.94
16.73
17.17
17.26
Mirror2-GS
17.20
17.60
17.98
17.04
17.09
QuerySplat
17.77
18.26
18.33
17.84
16.90
Appendix
Table 8: Quantitative comparison across seven benchmark datasets using 32 input views . The top-3 results are highlighted as first , second , and third .
Input
Method
N=4
N=8
N=16
N=32
N=64
w/o Pose
AnySplat
250
487
1,089
2,743
7,798
DA3-GS
709
1,513
3,783
10,950
35,904
Mirror2-GS
246
448
959
2,268
6,060
QuerySplat
2,324
3,398
6,092
14,009
39,583
Ours
411
761
1,649
3,991
11,037
w/ Pose
DepthSplat
158
287
541
1,045
2,087
Appendix
Table 11: Median forward-pass latency per scene (ms) of different methods under different numbers of input views. The fastest method within each pose setting is highlighted in bold.
Table 12: Ablation study of different normalization spaces across different numbers of input views.
Figure 11: Qualitative comparisons across diverse scenes and viewpoints with multi-view input.
Figure 12: Qualitative comparisons across diverse scenes and viewpoints with monocular input.
Figure 13: Qualitative results of SplatAnything for panoramic images reconstruction. SplatAnything can reconstruct high-quality 3DGS representations from panoramic images with different field of view, demonstrating its strong generalization ability.
Figure 14: More qualitative results of our method. SplatAnything feed-forward reconstructs high-quality 3DGS representations generalizing to arbitrary scenes, scales, and viewpoints in a few seconds.
Recent progress in feed-forward 3D Gaussian Splatting (3DGS) has notably improved rendering quality. However, the spatially uniform and highly redundant 3DGS map generated by previous feed-forward 3DGS methods limits their integration into downstream reconstruction tasks. We propose SparseSplat, the first feed-forward 3DGS model that adaptively adjusts Gaussian density according to scene structure and information richness of local regions, yielding highly compact 3DGS maps. To achieve this, we propose entropy-based probabilistic sampling, generating large, sparse Gaussians in textureless areas and assigning small, dense Gaussians to regions with rich information. Additionally, we designed a specialized point cloud network that efficiently encodes local context and decodes it into 3DGS attributes, addressing the receptive field mismatch between the general 3DGS optimization pipeline and feed-forward models. Extensive experimental results demonstrate that SparseSplat can achieve state-of-the-art rendering quality with only 22% of the Gaussians and maintain reasonable rendering quality with only 1.5% of the Gaussians. Project page: https://victkk.github.io/SparseSplat-page/.
In this work, we revisit several key design choices of modern Transformer-based approaches for feed-forward 3D Gaussian Splatting (3DGS) prediction. We argue that the common practice of regressing Gaussian means as depths along camera rays is suboptimal, and instead propose to directly regress 3D mean coordinates using only a self-supervised rendering loss. This formulation allows us to move from the standard encoder-only design to an encoder-decoder architecture with learnable Gaussian tokens, thereby unbinding the number of predicted primitives from input image resolution and number of views. Our resulting method, TokenGS, demonstrates improved robustness to pose noise and multiview inconsistencies, while naturally supporting efficient test-time optimization in token space without degrading learned priors. TokenGS achieves state-of-the-art feed-forward reconstruction performance on both static and dynamic scenes, producing more regularized geometry and more balanced 3DGS distribution, while seamlessly recovering emergent scene attributes such as static-dynamic decomposition and scene flow.
Jiawei Ren, Michal Jan Tyszkiewicz, Jiahui Huang +1
Recent feed-forward Gaussian reconstruction models adopt a pixel-aligned formulation that maps each 2D pixel to a 3D Gaussian, entangling Gaussian representations tightly with the input images. In this paper, we propose AnchorSplat, a novel feed-forward 3DGS framework for scene-level reconstruction that represents the scene directly in 3D space. AnchorSplat introduces an anchor-aligned Gaussian representation guided by 3D geometric priors (e.g., sparse point clouds, voxels, or RGB-D point clouds), enabling a more geometry-aware renderable 3D Gaussians that is independent of image resolution and number of views. This design substantially reduces the number of required Gaussians, improving computational efficiency while enhancing reconstruction fidelity. Beyond the anchor-aligned design, we utilize a Gaussian Refiner to adjust the intermediate Gaussiansy via merely a few forward passes. Experiments on the ScanNet++ v2 NVS benchmark demonstrate the SOTA performance, outperforming previous methods with more view-consistent and substantially fewer Gaussian primitives.