InfiniSplat: Implicit Gaussian Decoding for Large-Baseline Monocular View Synthesis

Under Review

Jiawei Wang^1* Hao Yu^1* Yongzhen Hu¹ Xinyi Yang¹ Tao Ni² Xin Zhan² Junbo Chen²

Xiaowei Zhou³ Ruizhen Hu⁴ Sida Peng^1†

¹Zhejiang University ²UDeer.ai ³State Key Laboratory of CAD&CG, Zhejiang University ⁴Shenzhen University

^*Equal contribution ^†Corresponding author

Paper

Code Coming Soon

TL;DR

InfiniSplat is a feed-forward single-image 3D Gaussian Splatting framework for large-baseline novel view synthesis.
It moves Gaussian generation from pixel-aligned image grids toward surface-aligned supports guided by monocular geometry.
A query-conditioned implicit Gaussian decoder predicts Gaussian attributes on sampled supports, producing more coherent surfaces under large viewpoint changes.

Abstract

Single-image feed-forward 3D Gaussian Splatting (3DGS) aims to directly generate a renderable 3D scene representation from one input image, avoiding the cost of multi-view capture and per-scene optimization. However, existing methods are often constrained by a pixel-aligned representation, where Gaussians are predicted from fixed image-grid locations. Such pixel-aligned primitives can produce promising nearby-view renderings, but they remain weakly coupled to underlying scene surfaces and struggle to preserve coherent structures under large viewpoint shifts. We present InfiniSplat, a feed-forward single-image 3DGS framework that moves from a pixel-aligned representation toward a surface-aligned representation. InfiniSplat constructs this representation by first using geometry-guided sampling to place 2D supports according to depth-induced local surface structure, and then applying a query-conditioned implicit decoder to predict Gaussian attributes from the image features queried at these supports. By grounding support locations in geometry while decoupling Gaussian prediction from fixed pixel centers, InfiniSplat produces Gaussian layouts that better follow scene surfaces and reduce scattered primitives caused by grid discretization. Across multiple cross-dataset NVS evaluations, InfiniSplat achieves state-of-the-art performance compared with single-image feed-forward baselines, and demonstrates zero-shot generalization from Hypersim indoor synthetic training to complex open-world scenes.

Method

Pipeline of InfiniSplat:

Geometry-guided support sampling: InfiniSplat predicts geometry priors from a single input image and samples 2D supports according to depth-induced local surface structure rather than fixed image-grid locations.
Implicit Gaussian decoding: the decoder queries image features at sampled supports and predicts bounded updates to produce a surface-aligned 3D Gaussian representation for feed-forward novel view synthesis.

Qualitative Comparison

RGB qualitative comparison under large viewpoint changes. InfiniSplat-RGB produces fewer visible cracks and more coherent scene structures than feed-forward baselines.

With sparse LiDAR points as additional input, InfiniSplat-LiDAR produces substantially better novel-view rendering quality than ADGaussian.

Citation

@article{wang2026infinisplat,
    title={InfiniSplat: Implicit Gaussian Decoding for Large-Baseline Monocular View Synthesis},
    author={Wang, Jiawei and Yu, Hao and Hu, Yongzhen and Yang, Xinyi and Ni, Tao and Zhan, Xin and Chen, Junbo and Zhou, Xiaowei and Hu, Ruizhen and Peng, Sida},
    journal={arXiv preprint arXiv:xxxx.xxxxx},
    year={2026}
}