[CVPRF 2026] LangFlash: Feed-forward 3D Language Gaussian Splatting from Sparse Unposed Images

Yilong Liu, Wanhua Li, Chen Zhu-Tian, Hanspeter Pfister

[Paper] [Project Page]

LangFlash is a feed-forward framework for 3D language Gaussian splatting from sparse, unposed input images.

Installation

Create a Python 3.10 environment and install the required dependencies (other torch version should work):

conda create -y -n langflash python=3.10
conda activate langflash
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126
pip install -r requirements.txt

Pre-trained Checkpoints

Download the geometry and semantic checkpoints, then place both files in ckpt/:

• Geometry checkpoint: ckpt/mixRe10kDl3dv.ckpt （From NoPoSplat)
• Semantic checkpoint: ckpt/re10k.ckpt

The final directory layout should look like this:

ckpt/
  mixRe10kDl3dv.ckpt
  re10k.ckpt

Demo

Run the Gradio demo with:

python -m LangFlash.model.gradio

The demo uses the bundled examples under sample/ and loads checkpoints from ckpt/.

Dataset

The preprocessed dataset is available on Hugging Face.

The dataset release contains:

scannet_extracted_sp   ScanNet data and metadata
clip_test              RE10K test language features and corresponding masks
clip_train_part{i}      RE10K training language features and corresponding masks

Acknowledgements

This project builds on several excellent open-source projects:NoPoSplat,EfficientViT,LangSplat,SAM2,lseg-minimal, and FG-CLIP.

Citation

If you find LangFlash useful, please cite:

@misc{liu2026langflash,
  title         = {LangFlash: Feed-forward 3D Language Gaussian Splatting from Sparse Unposed Images},
  author        = {Yilong Liu and Wanhua Li and Chen Zhu-Tian and Hanspeter Pfister},
  year          = {2026},
  eprint        = {2605.23287},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}