Official code release for the paper: TSGS: Improving Gaussian Splatting for Transparent Surface Reconstruction via Normal and De-lighting Priors.

Mingwei Li^1,2, Pu Pang^3,2, Hehe Fan¹, Hua Huang⁴, Yi Yang^1,✉

¹Zhejiang University, ²Zhongguancun Academy, Beijing, ³Xi'an Jiaotong University, ⁴Beijing Normal University

• [2026-03-19]: v1.3 released! — Fix critical K-matrix scaling bug for DTU at resolution > 1, enhance DTU pipeline with full CLI control. With Lotus-2 normal priors + ASG, mean CD on DTU improves from 0.511 (paper) to 0.507.
• [2026-03-14]: v1.2 released — Support TransNormal as an alternative normal prior via --normal_folder, achieving improved geometric reconstruction (avg CD 1.68 vs 1.85).
• [2026-03-14]: v1.1 released — Optimized hyperparameters, parameterized CUDA rasterizer thresholds, fixed rendering pipeline (AppModel/SpecularModel loading), and configurable random seed.
• [2025-07-13]: 🎉 Our code and dataset are released!
• [2025-07-05]: 🏆 Our paper has been accepted by ACM MM 2025!
• [2025-04-18]: 🎉 Our arXiv paper is released! You can find it here. Project page is also live!

We present TSGS, a framework for high-fidelity transparent surface reconstruction from multi-views. (a) We introduce TransLab, a novel dataset for evaluating transparent object reconstruction. (b) Comparative results on TransLab demonstrate the superior capability of TSGS.

(a) The two-stage training process. Stage 1 optimizes 3D Gaussians using geometric priors and de-lighted inputs. Stage 2 refines appearance while fixing opacity. (b) Inference extracts the first-surface depth map for mesh reconstruction. (c) The first-surface depth extraction module uses a sliding window for robust depth calculation.

1. Clone the repository and setup environment:

   git clone https://github.com/longxiang-ai/TSGS.git cd TSGS conda create -n tsgs python=3.8 -y # Tested with Python 3.8, other versions may also work conda activate tsgs

2. Install dependencies: Install PyTorch matching your CUDA version (see PyTorch website for the correct command). We have tested with Python 3.8 and CUDA 11.8, but other corresponding PyTorch-CUDA versions should also work. Example for CUDA 11.8:

   pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118 # Install other requirements and submodules pip install -r requirements.txt pip install submodules/diff-first-surface-rasterization pip install submodules/simple-knn

3. Install StableNormal (for input preprocessing): If you need to generate normal and de-lighting maps as input priors, install the StableNormal repository:

   git clone https://github.com/Stable-X/StableNormal.git cd StableNormal pip install -r requirements.txt python preprocess/process_normal.py --source_path /path/to/your/data # For our provided TransLab dataset, you can skip this step because the normal and de-lighting maps are already provided. cd .. # Return to the TSGS directory

4. (Optional) Install TransNormal (alternative normal prior, v1.2): TransNormal can be used as an alternative normal estimation model, which yields improved geometric reconstruction quality.

   git clone https://github.com/longxiang-ai/TransNormal.git # Follow TransNormal's README to download model weights # Generate normal maps for all scenes: python preprocess/process_transnormal.py --all --data_root data/translab \ --transnormal_root /path/to/TransNormal # Then train with: --normal_folder transnormals

We introduce TransLab, a novel dataset specifically designed for evaluating transparent object reconstruction in laboratory settings. It features 8 diverse, high-resolution 360° scenes with challenging transparent glassware. Details of collecting the dataset can be found in Translab.

Download options:

• HuggingFace (recommended):

  # Install huggingface_hub if needed: pip install huggingface_hub huggingface-cli download Longxiang-ai/TransLab --repo-type dataset --local-dir data/translab

• Google Drive: Download link

Please put downloaded data in the data folder, and the structure should be like this:

data/ ├── translab/ │ ├── scene_01/ │ │ ├── images/ # original images RGB channel, mask as A channel │ │ ├── masks/ # Rendered by Blender │ │ ├── normals/ # obtained by StableNormal │ │ ├── transnormals/ # (optional) obtained by TransNormal (v1.2) │ │ ├── delights/ # obtained by StableDelight │ │ ├── sparse/ # obtained by colmap │ │ ├── meshes/ # exported from blender, for mesh evaluation │ │ └── transparent_masks/ # rendered by Blender │ ├── scene_02/ │ ├── ... │ └── scene_08/ ├── dtu_dataset/

We follow the same data preparation process as PGSR to prepare the DTU dataset.

Put dtu data in the data folder, and the structure should be like this:

data/ ├── dtu_dataset/ │ ├── dtu/ │ │ ├── scan24/ │ │ ├── ... │ ├── dtu_eval/ │ │ ├── ObsMask/ │ │ ├── Points

The following scripts will first train each scene in the dataset, and then evaluate the results.

sh run_translab.sh # run on TransLab dataset sh run_dtu.sh # run on DTU dataset

v1.2 (TransNormal prior):

# First generate TransNormal normal maps (see Installation step 4) # Then train with the alternative normal folder: python scripts/run_translab.py --normal -d --eval --out_name tsgs_transnormal \ --normal_folder transnormals --seed 7

v1.3 (DTU with Lotus-2 normal prior):

v1.3 fixes a critical K-matrix scaling bug when using --resolution 2 on DTU. Normal priors are generated by Lotus-2. The full_paper config (Lotus-2 normals + ASG + SD guidance) is now the default for run_dtu.sh.

# 1. Generate Lotus-2 normal maps for all DTU scans python scripts/generate_dtu_normals.py --scan_id 24 --gpu_id 0 # repeat for each scan # 2. Run DTU (all params already set as defaults) sh run_dtu.sh

Key observations:

• v1.1 (optimized hyperparameters + StableNormal): Improves over paper baselines in most image quality metrics. Average PSNR improves from 39.08 to 39.31.
• v1.2 (TransNormal prior): Achieves significant geometric improvement — average Chamfer Distance drops from 1.86 (paper) / 1.81 (v1.1) to 1.68, and F1 score improves from 0.955 to 0.961, while maintaining comparable image quality.

v1.3 uses Lotus-2 normal priors + Adaptive Spherical Gaussians (ASG) + Score Distillation Normal Guidance, with a single consistent config across all 15 scans. Key bug fix: K-matrix scaling for NonCenteredCamera at resolution > 1.

Key observations:

• v1.3 achieves mean CD 0.507, improving over the paper's 0.511 with a single consistent hyperparameter setting.
• 9/15 scans match or beat the paper's reported values, with notable improvements on scan55 (-0.087), scan110 (-0.038), and scan65 (-0.025).

• Release Arxiv paper link.
• Release source code.
• Release TransLab-Synthetic dataset and download link.
• v1.1: Optimized hyperparameters and rendering pipeline fixes.
• v1.2: TransNormal support for improved geometric reconstruction.
• v1.3: Fix K-matrix scaling bug, DTU pipeline with Lotus-2 normals (mean CD 0.507 vs paper 0.511).
• Release TransLab-Real dataset and download link.
• Provide detailed installation and usage instructions.

We would like to thank the following open-source projects for their valuable contributions: PGSR, StableNormal, TransNormal, Lotus-2, 2DGS, and GroundedSAM.

We also thank Nerfies for their amazing website template.

If you find our work useful, please consider citing:

@misc{li2025tsgs, title={TSGS: Improving Gaussian Splatting for Transparent Surface Reconstruction via Normal and De-lighting Priors}, author={Mingwei Li and Pu Pang and Hehe Fan and Hua Huang and Yi Yang}, year={2025}, eprint={2504.12799}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2504.12799}, }