GeoGSplat. Paper

FEED-FORWARD 3D RECONSTRUCTION & NOVEL VIEW SYNTHESIS

GeoGSplat
Geometry-Grounded
Non-Pixel-Aligned Gaussian Splatting

Jiale Wang1,3*Chen Zhao2*Haoyang Zhang3Yanhao Wu1,3Fei He3Yanhu Shan3Rui Wu3Wei Ke1Tong Zhang4‡
1 Xi’an Jiaotong University2 EPFL3 Horizon Robotics4 University of the Chinese Academy of Sciences

* Equal contribution   ·   ‡ Corresponding author

PaperCode Coming soon
Figure 1: input images, TokenGS and GeoGSplat reconstructions and novel-view renderings, alongside a comparison of reconstruction quality, rendering quality, and Gaussian count.

FIGURE 1 / OVERVIEW GeoGSplat combines Anchor and Triplane representations for 3D reconstruction and novel view synthesis. Larger bubbles indicate more Gaussians.

Original figure ↗

OVERVIEW

Abstract

Feed-forward Gaussian Splatting enables efficient novel view synthesis. Pixel-aligned prediction makes the primitive count grow with image resolution and view count, introducing redundancy. Non-pixel-aligned methods predict a fixed number of Gaussians, but without explicit geometric constraints, misplaced centers can cause rendering artifacts under large viewpoint changes.

GeoGSplat incorporates geometric priors into non-pixel-aligned prediction. A dual-branch structured encoder constructs Anchors and Triplane to capture coarse 3D structure and fine-grained scene features. A geometry-guided hierarchical decoder progressively predicts dense 3D points and Gaussian primitives. Experiments on DTU and NRGBD show improvements in novel view synthesis and 3D reconstruction, with a larger advantage over TokenGS as target viewpoints move farther from the input cameras.

Dual-branch encoder and hierarchical decoder
65,536Predicted Gaussians, independent of input-view count
20.48 dBDTU PSNR with four input views · TokenGS: 17.71 dB
0.017NRGBD Chamfer distance · TokenGS: 0.277 · Four input views

3D RECONSTRUCTION

Point-cloud comparison

GeoGSplat and original TokenGS Gaussian centers.
Both models use the same context views.

Reconstruction examples

Select a scene to update the comparison below.

Laughing BuddhaObject-level
65,536 centers per method
Point-cloud preview GeoGSplat Loading geometry…
DRAG TO ORBIT · SCROLL TO ZOOM · SHIFT + DRAG TO PAN
TokenGS Gaussian-center preview TokenGS Loading geometry…
SYNCHRONIZED CAMERA
Context images

Same context views and coordinate system for both methods.

All 65,536 Gaussian centers per method, in a shared coordinate system.

NOVEL VIEW SYNTHESIS

Qualitative comparison

Figure 4: novel view synthesis
with four input views.

01 / OBJECT-LEVELObject detail
Object-level novel-view comparisons from paper Figure 4: GT, AnySplat with ground-truth poses, TokenGS, and GeoGSplat. Red boxes highlight object markings.
02 / SCENE-LEVELScene structure
Scene-level novel-view comparisons from paper Figure 4: GT, AnySplat with ground-truth poses, TokenGS, and GeoGSplat. Red boxes highlight kitchen objects and chair geometry.

Columns show ground truth, AnySplat*, TokenGS, and GeoGSplat. Red boxes highlight blurred details in TokenGS and shifted or duplicated structures in AnySplat*. GeoGSplat preserves finer details and more accurate scene structures in these regions. AnySplat* uses ground-truth poses. Original images and annotations are retained from the paper.

Original Figure 4 ↗

METHOD

GeoGSplat architecture

Original Figure 3 ↗
GeoGSplat architecture: images and Plücker rays feed a dual-branch Anchor and Triplane encoder, followed by Anchor-to-Point and Point-to-Gaussian decoding
01 / REPRESENTATIONS

Dual-Branch Structured Encoder

Image and Plücker ray embeddings go through a Vision Transformer. Shared attention turns image features into Anchors (learnable centers and radii) and a Triplane of fine-grained scene features.

02 / GEOMETRY

Anchor-to-Point Decoding

Reference points start inside each Anchor region. Triplane features and cross-attention predict offsets that move them into dense 3D points. Training uses Chamfer distance to align Anchor centers and dense points with the ground truth.

03 / GAUSSIAN PRIMITIVES

Point-to-Gaussian Decoding

The decoder queries the Triplane at these points and predicts Gaussian parameters with a shared MLP head. 4,096 Anchors × 16 points gives 65,536 Gaussians, independent of input-view count and image resolution.

EXPERIMENTS

Main results

Novel view synthesis and 3D reconstruction.
Four input views, as in the main tables.

↑ Higher is better   ·   ↓ Lower is better
MethodPSNR ↑SSIM ↑LPIPS ↓CD ↓Gaussians

AnySplat* uses ground-truth poses. CD is the average of Pred→GT and GT→Pred Chamfer distances in a shared normalized coordinate system. GeoGSplat is evaluated on dense points; other methods use Gaussian centers. The appendix reports nearly identical GeoGSplat results using Gaussian centers.

VIEWPOINT-DISTANCE ANALYSIS

Novel view synthesis under
larger viewpoint changes

GeoGSplat and TokenGS are compared along the same target-view trajectories with fixed context views. Performance is similar near the context cameras, and GeoGSplat’s PSNR advantage grows with target-context distance.

Target-context distance is the distance from the target camera center to the nearest context camera center. PSNR is computed on target pixels visible in at least one context view.

Figure 5: PSNR and the GeoGSplat advantage over TokenGS at increasing target-context distances

PREPRINT

Paper and citation

The paper is available on arXiv.
If you find this work useful, please cite:

Paper
BIBTEX
@misc{wang2026geogsplat,
  title         = {GeoGSplat: Geometry-Grounded
                   Non-Pixel-Aligned Gaussian Splatting},
  author        = {Jiale Wang and Chen Zhao and Haoyang Zhang and
                   Yanhao Wu and Fei He and Yanhu Shan and
                   Rui Wu and Wei Ke and Tong Zhang},
  year          = {2026},
  eprint        = {2610.36760},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2610.36760}
}