Sparse-to-Complete: From Sparse Image Captures to Complete 3D Scenes
Yiyang Shen, Yin Yang, Kun Zhou, Tianjia Shao
SIGGRAPH 2026

[Project Page] [Paper]

This repository contains the source code for "Sparse-to-Complete: From Sparse Image Captures to Complete 3D Scenes".

Clone the repository with the Pi3 submodule:

git clone --recursive https://github.com/yyyoungs/S2C-3D.git
cd S2C-3D

For an existing clone, initialize the submodule with:

git submodule update --init --recursive

Create the S2C-3D conda environment:

conda env create -f environment.yml
conda activate S2C-3D

The default environment uses PyTorch 2.4.0 with CUDA 12.1 wheels. If your CUDA driver or platform requires another PyTorch build, install the matching PyTorch wheel first and then install the remaining packages from environment.yml.

If the CUDA renderer extension is not built in your environment, build it with:

cd cuda_renderer
python setup.py install
cd ..

Each scene should contain an images/ folder. The following tree shows the bundled example scene:

example/6_views/
└── 52599ae063/
    └── images/
        ├── inputs_0000.png
        ├── inputs_0001.png
        ├── inputs_0002.png
        └── ...

Image names are sorted lexicographically. Numbered names such as inputs_0000.png, inputs_0001.png, ... are recommended.

For your own data, replace 52599ae063 with your scene name and put all input images under that scene's images/ folder.

Pi3 preparation is run automatically by main.py. It writes:

scene_id/
├── images/
└── all_views/
    ├── cam_idx.json
    ├── sparse/
    │   ├── cameras.bin
    │   ├── images.bin
    │   ├── points3D.bin
    │   └── points.ply
    └── train_img/

main_model.steps controls which stages run:

main_model:
  # [Pi3, phase1, phase2, phase3, gui]
  steps: [1, 1, 1, 1, 1]

Run all scenes under main_model.data_dir in configs/config.yaml:

conda activate S2C-3D
python main.py --cuda_devices 0

Run a single scene:

python main.py --cuda_devices 0 \
  main_model.case_dir=/path/to/scene

Outputs are written to:

<result_dir>/<scene_id>/
├── phase1/
├── phase2/
└── phase3/

The final Gaussian checkpoint is:

<result_dir>/<scene_id>/phase3/gs/gs.pth

For example, after a scene is fully reconstructed, open only the GUI:

python main.py --cuda_devices 0 \
  main_model.case_dir=/path/to/scene \
  'main_model.steps=[0,0,0,0,1]'

Pi3 preparation is normally called by main.py. You can also run it separately before launching the full pipeline:

python Pi3/pi3_prepare/batch_prepare.py \
  --scene_path /path/to/scene \
  --cuda_devices 0

Batch mode:

python Pi3/pi3_prepare/batch_prepare.py \
  --data_root /path/to/dataset \
  --view_dirs 6_views \
  --cuda_devices 0

Open the GUI for one reconstructed scene:

python main.py --cuda_devices 0 \
  main_model.case_dir=/path/to/scene \
  'main_model.steps=[0,0,0,0,1]'

For the bundled example:

python main.py --cuda_devices 0 \
  main_model.case_dir=./example/6_views/52599ae063 \
  'main_model.steps=[0,0,0,0,1]'

Basic controls:

• W/S: move forward/backward.
• A/D: move left/right.
• Q/E: move down/up.
• J/L: yaw left/right.
• I/K: pitch up/down.
• U/O: roll left/right.
• Mouse wheel: zoom.
• Left drag: orbit.
• Middle drag: pan.

S2C-3D/
├── main.py
├── configs/
├── scripts/
│   ├── plan_cameras.py
│   ├── prepare_gaussians.py
│   ├── train_lora.py
│   ├── refine_scene.py
│   └── view_scene.py
├── model/
├── Pi3/
│   ├── Pi3-main/
│   └── pi3_prepare/
├── cuda_renderer/
└── lpipsPyTorch/

@article{shen2026sparse,
  title={Sparse-to-Complete: From Sparse Image Captures to Complete 3D Scenes},
  author={Shen, Yiyang and Yang, Yin and Zhou, Kun and Shao, Tianjia},
  journal={arXiv preprint arXiv:2605.05664},
  year={2026}
}

This project builds on excellent open-source work including 3DGS, gsplat, Pi3, and Difix3D+. We thank all the authors for their great work.