ExploreGS

ExploreGS: Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors (ICCV 2025)

---

👥 Authors

Minsu Kim, Subin Jeon, In Cho, Mijin Yoo, Seon Joo Kim

🔗 Links

Project page | Paper | Dataset

---

🗞️ News

• [05/18/2026] Dataset generation pipeline for diffusion training is now available!
• [10/31/2025] Code and models are now available!

---

⚙️ Setup

1. Code setup

   git clone https://github.com/minsu1206/ExploreGS.git --recursive
   cd ExploreGS
   conda env create -f environment.yml
   conda activate exploregs
   pip install submodules/simple-knn
   pip install submodules/upgrade-diff-gaussian-rasterization
   pip install submodules/diff-gaussian-rasterization-fisherrf (If you want ablation)
   

   Instead of conda install, you can download docker image

   docker pull minsu1206/exploregs
   

2. Download dataset & checkpoint

   • Run download.sh
   • Edit dataset & checkpoint path at run_stage1.sh , run_stage2.sh , run_stage1_we.sh, run_stage2_we.sh
   • Download Depth-Anything-V2 checkpoint. we use vitl version.
   • If you want to preprocess for custom dataset, read README.md at folder data_preprocess (#TBD)

Optional: Dataset Generation

To create a dataset for training the diffusion model, refer to ExploreGS-GS-DataGen.

🚀 Run

▶ Curated Nerfbusters

bash scripts/commands_curatednb.sh

▶ WildExplore

bash scripts/commands_wildexplore.sh

---

Acknowledgements

Our work is built upon these excellent open-source projects:

• 3DGS
• Nerfbusters
• DL3DV-10K
• 3DGS-Enhancer
• DynamiCrafter

Citation

If you find this work useful, please consider citing:

@misc{kim2025exploregsexplorable3dscene,
      title={ExploreGS: Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors}, 
      author={Minsu Kim and Subin Jeon and In Cho and Mijin Yoo and Seon Joo Kim},
      year={2025},
      eprint={2508.06014},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2508.06014}, 
}