GLINT: Modeling Scene-Scale Transparency via Gaussian Radiance Transport

Official code release for the paper: GLINT: Modeling Scene-Scale Transparency via Gaussian Radiance Transport.

Youngju Na^1,2,*, Jaeseong Yun², Soohyun Ryu², Hyunsu Kim², Sung-Eui Yoon¹, Suyong Yeon²

¹KAIST, ²NAVER LABS

News

• [2026-04-09]: 🎉 Our paper has been selected as an Oral presentation at CVPR 2026.
• [2026-03-30]: Initial code release.

Overview

GLINT is a method for modeling large-scale transparent and reflective scenes with Gaussian radiance transport.

Installation

1. Clone the repository and setup environment:

conda create -n glint python=3.11 -y
conda activate glint

2. Install PyTorch:

Install PyTorch matching your CUDA version (see PyTorch website for the correct command). Example for CUDA 11.8:

pip install torch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 --index-url https://download.pytorch.org/whl/cu118

3. Install dependencies:

cat requirements.txt | sed -e '/^\s*-.*$/d' -e '/^\s*#.*$/d' -e '/^\s*$/d' | \
  awk '{split($0, a, "#"); if (length(a) > 1) print a[1]; else print $0;}' | \
  awk '{split($0, a, "@"); if (length(a) > 1) print a[2]; else print $0;}' | \
  xargs -n 1 pip install

pip install -e . --no-build-isolation --no-deps

4. Install submodules:

git submodule update --init --recursive
pip install -v submodules/diff-surfel-tracing
pip install \
  submodules/diff-surfel-rasterizations/diff-surfel-rasterization-wet \
  submodules/diff-surfel-rasterizations/diff-surfel-rasterization-wet-ch06 \
  submodules/diff-surfel-rasterizations/diff-surfel-rasterization-wet-ch08

Data Preparation

The ref-dl3dv and 3D-FRONT-T dataset used in our paper is available for download:

Download Dataset (Google Drive)

GLINT expects datasets in the EasyVolcap-style format. At minimum, each scene should provide:

<scene>/
├── images/
├── intri.yml
├── extri.yml
└── sparse/

For the G-buffer guidance, each scene also contains priors obtained from DiffusionRenderer. You may also consider using other useful priors (e.g., TransNormal, Video Depth Anything, etc.). Please refer to these if you want to build your custom datasets:

<scene>/
├── images/
├── intri.yml
├── extri.yml
├── sparse/
├── envs/
│   └── points3D.ply
├── normals/
│   └── <view_id>/000000.jpg
└── diffrens/
    ├── normal/<view_id>/000000.png
    ├── depth/<view_id>/000000.png
    ├── diffuse_albedo/<view_id>/000000.png
    ├── basecolor/<view_id>/000000.png
    ├── roughness/<view_id>/000000.png
    └── metallic/<view_id>/000000.png

The diffusion-renderer prior maps must be placed under diffrens/.

For training and evaluation, <scene> should match the dataset directory name and the config filename in configs/exps/glint/ref-dl3dv/.

Scene list used in our dl3dv-10k subset (ref-dl3dv).

194defaa605986166d52ae703b1d44d1a557794698386becaaa5f688f4fb026b
3712b8fdcb94128c92c2e2c30fb529851e3231cdc7c4451bc6c784f923386e93
52410f0264d14bde6acd695c637aaa274833be8afcf05ef4fd6a51176ad2dbd2
543b6607de9318e3a0c68b267a4b616fdc5849a140ba184807d5e70e567f8ec0
5454b71d612cc2b020e60bd2d8a018dc33d62b5fbd5c041b55a752480a8a97ba
6b42314a2f8a18a193826e2b58e45729453e74524078283f740b8f8d330c3d2f
b65e86833c1ae29714ce881bb9d14d3ed1256a08ab944fd9e75d6b29c674346d
b9df30d6e6078880acc88acb01872c65d337f84b9dba44a23fa29c9861d7e23b

If you want to prepare your own data from COLMAP outputs, see the preprocessing scripts in scripts/preprocess/.

Training

Example training command:

evc-train -c configs/exps/glint/ref-dl3dv/<scene>.yaml \
  exp_name=glint/ref-dl3dv/<run_name>/<scene>

For example, to train the scene 6b42314a2f8a18a193826e2b58e45729453e74524078283f740b8f8d330c3d2f:

evc-train -c configs/exps/glint/ref-dl3dv/6b42314a2f8a18a193826e2b58e45729453e74524078283f740b8f8d330c3d2f.yaml \
  exp_name=glint/ref-dl3dv/<run_name>/6b42314a2f8a18a193826e2b58e45729453e74524078283f740b8f8d330c3d2f

In this scene config, training uses dataloader_cfg.dataset_cfg.view_sample, which contains all view indices except multiples of 8.

The default hyperparameters are defined in configs/models/glint.yaml. Key parameters you may want to adjust depending on your scene:

Parameter	Default	Description
render_reflection_start_iter	3000	Iteration to start reflection rendering
render_transmission_start_iter	1000	Iteration to start transmission rendering
depth_discrepancy_threshold	0.005	Depth discrepancy threshold (scene-scale dependent)
trans_map_reg_loss_weight	0.01	Transmission map regularization weight
trans_guidance_loss_weight	0.01	Transmission guidance loss weight

Evaluation

Example evaluation command:

evc-test -c configs/exps/glint/ref-dl3dv/<scene>.yaml \
  exp_name=glint/ref-dl3dv/<run_name>/<scene>

For the same scene:

evc-test -c configs/exps/glint/ref-dl3dv/6b42314a2f8a18a193826e2b58e45729453e74524078283f740b8f8d330c3d2f.yaml \
  exp_name=glint/ref-dl3dv/<run_name>/6b42314a2f8a18a193826e2b58e45729453e74524078283f740b8f8d330c3d2f

Evaluation uses val_dataloader_cfg.dataset_cfg.view_sample. For 6b42314a2f8a18a193826e2b58e45729453e74524078283f740b8f8d330c3d2f, this is:

[0, 8, 16, 24, ..., 320]

So this scene follows an every-8th-view evaluation split: 41 evaluation views and the remaining 282 views for training.

Custom Rendering

Example interpolation video rendering:

bash scripts/render_interp_video.sh \
  --config configs/exps/glint/ref-dl3dv/<scene>.yaml \
  --exp_name glint/ref-dl3dv/<run_name>/<scene> \
  --cam_idx1 0 \
  --cam_idx2 8 \
  --n_frames 60

Repository Structure

• easyvolcap/ — Core framework and GLINT model implementation
• configs/ — Model, dataset, and experiment configurations
• scripts/ — Preprocessing, training utilities, and rendering scripts
• submodules/ — Required custom CUDA and tracing dependencies

Roadmap

• Release source code.
• Release 3D-FRONT-T Blender files for downstream applications.

Acknowledgements

This codebase is built on top of EasyVolcap and the 2D Gaussian ray tracer from EnvGS. We sincerely thank the authors and contributors of these projects. You may also want to check out the related works listed below.

Related Work

• TSGS: Improving Gaussian Splatting for Transparent Surface Reconstruction via Normal and De-lighting Priors
• TransparentGS: Fast Inverse Rendering of Transparent Objects with Gaussians
• DiffusionRenderer: Neural Inverse and Forward Rendering with Video Diffusion Models
• TransNormal: Dense Visual Semantics for Diffusion-based Transparent Object Normal Estimation

Citation

If you find this repository useful, please consider citing our paper:

@misc{na2026glint,
  title={GLINT: Modeling Scene-Scale Transparency via Gaussian Radiance Transport},
  author={Youngju Na and Jaeseong Yun and Soohyun Ryu and Hyunsu Kim and Sung-Eui Yoon and Suyong Yeon},
  year={2026},
  eprint={2603.26181},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2603.26181},
}

License

This project is released under the MIT License.