TL;DR
- InfiniSplat v2 reconstructs 3D Gaussian scenes from arbitrary input views, with or without camera poses.
- A dual-branch architecture freezes geometry and learns Gaussian attributes, with opacity suppressing artifacts from inconsistent points.
- Training with scale-aligned ground-truth cameras and large-angle target views improves rendering under large viewpoint changes.
Demo
Visualization
Drag to orbit · Scroll to zoomFly mode: WASD + Mouse · Esc to exit
Comparison between Reconstructed Gaussians and Point Clouds
Point CloudsGaussians
Loading comparison…
Drag to compare.
Comparison
Large-scale Reconstruction
Large-scale feed-forward reconstruction on KITTI · 1,000+ frames.
Acknowledgements
We are grateful to Yujie Guo for insightful discussions on experimental analysis and the interpretation of our results. These exchanges helped us examine our findings more critically and improve the clarity of our experimental discussion.
Citation
@article{hu2026infinisplatv2,
title={InfiniSplat v2: Analyzing and Improving Feed-Forward 3DGS from Any Views},
author={Hu, Yongzhen and Yu, Hao and Wang, Jiawei and Zeng, Xianming and Xu, Jianyun and Liu, Xingliang and Hu, Ruizhen and Zhou, Xiaowei and Bao, Hujun and Peng, Sida},
journal={arXiv preprint},
eprint={},
year={2026}
}











































