SparseSplat: Towards Applicable Feed-Forward 3D Gaussian Splatting with Pixel-Unaligned Prediction

April 8, 2026 · View on GitHub

Project Page | Paper (PDF) | Code (Coming Soon)

CVPR 2026

Zicheng Zhang¹, Xiangting Meng², Ke Wu¹, Wenchao Ding²

¹Fudan University ²ShanghaiTech University

Abstract

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 3D-Local Attribute Predictor 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.

Key Results (DL3DV)

Method	PSNR ↑	SSIM ↑	LPIPS ↓	GS Count ↓
MVSplat	22.95	0.774	0.192	688k
DepthSplat	24.17	0.816	0.152	688k
SparseSplat (Ours, 150k)	24.20	0.817	0.168	150k
SparseSplat (Ours, 40k)	22.65	0.737	0.251	40k
SparseSplat (Ours, 10k)	21.29	0.665	0.321	10k

SparseSplat matches SOTA rendering quality (DepthSplat) using only 22% of the Gaussians, and remains faster than DepthSplat at the 10k and 40k settings — ideal for real-time SLAM and edge applications.

Code

The training and inference code will be released here soon. The code/ directory is reserved as a placeholder for the upcoming release.

BibTeX

@inproceedings{zhang2026sparssplat,
  title={SparseSplat: Towards Applicable Feed-Forward 3D Gaussian Splatting with Pixel-Unaligned Prediction},
  author={Zhang, Zicheng and Meng, Xiangting and Wu, Ke and Ding, Wenchao},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026}
}

Website License

This project page was built using the Academic Project Page Template. Licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.