[CVPR 2026] LightSplat: Fast and Memory-Efficient
Open-Vocabulary 3D Scene Understanding in Five Seconds

Jaehun Bang¹ · Jinhyeok Kim^2* · Minji Kim¹ · Seungheon Jeong¹ · Kyungdon Joo^1†
1AIGS, UNIST  ·  ²GSAI, POSTECH
^*Work done while at UNIST  ·  ^†Corresponding author

Overview

LightSplat is a fast and memory-efficient training-free framework for open-vocabulary 3D scene understanding. It injects compact 2-byte semantic indices into 3D Gaussian representations from multi-view images, avoiding iterative feature optimization and dense per-Gaussian feature storage.

Dataset Preparation

 Setup guide       Dataset files

DL3DV-OVS is the four-scene benchmark introduced with LightSplat for open-vocabulary understanding of large, complex indoor and outdoor scenes. The release includes GT masks, SAM/CLIP features, text embeddings, and RGB 3DGS checkpoints.

Code

Code will be released soon.

Citation

@inproceedings{bang2026lightsplat,
  title={Lightsplat: Fast and memory-efficient open-vocabulary 3d scene understanding in five seconds},
  author={Bang, Jaehun and Kim, Jinhyeok and Kim, Minji and Jeong, Seungheon and Joo, Kyungdon},
  booktitle={2026 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages={19812--19821},
  year={2026},
  organization={IEEE}
}