Generalizable Human Gaussian Splatting
via Multi-View Semantic Consistency

Official PyTorch implementation of
"Generalizable Human Gaussian Splatting via Multi-View Semantic Consistency"

Jingi Kim, Wonjun Kim (Corresponding Author)

⛰️CVPR 2026 Findings⛰️

Qualitative results of SemanticGHGS.

Overall architecture of SemanticGHGS.

---

👀 Overview

SemanticGHGS is a feed-forward framework for generalizable human Gaussian splatting from sparse multi-view inputs.
Our method is built on the key idea of multi-view semantic consistency, which encourages semantically corresponding regions across views to interact in a coherent 3D-aware manner.

Given sparse reference views, SemanticGHGS extracts multi-view visual features, lifts them into 3D-aware representations, and predicts Gaussian attributes for human rendering. By enforcing semantic consistency across views during feature aggregation, our framework improves cross-view correspondence, stabilizes Gaussian prediction, and produces more reliable renderings under sparse-view settings.

Highlights

• 🚀 A feed-forward framework for sparse-view human Gaussian splatting
• 🌍 Multi-view semantic consistency for robust cross-view feature fusion
• 🧠 3D-aware feature aggregation for improved Gaussian prediction
• 🎯 Generalizable rendering without per-subject optimization

---

📦 Installation

1) Clone

git clone https://github.com/jingi0614/SemanticGHGS_release.git
cd SemanticGHGS_release

2) Create environment + install dependencies

We provide an installer:

bash install.sh

This will:

• create a conda env (SemanticGHGS, Python 3.8)
• install PyTorch 2.0.1 + CUDA 11.8, PyTorch3D, iopath
• install Python deps via requirements.txt
• build/install ./submodules/diff-gaussian-rasterization/

---

🗂️ Dataset

Dataset Preparation

Please request and download the official THuman2.0 dataset.
The request form is available here: THuman2.0 Agreement Form

Recommended directory structure:

datasets/
└── THuman/
    ├── THuman2.0_Release/
    └── THuman2.0_smplx/

After downloading the raw scans, RGB images can be rendered by running:

python ./prepare_data/render_data.py

The corresponding SMPL-X parameters should be placed under the dataset directory.

🧪 Evaluation

Set the dataset path in the config file:

• config/config_thu.yaml

Download the pretrained checkpoint from link at here (THuman2.0, THuman2.1, 2K2K, Xhuman).

Evaluate the trained model with:

python test.py --config config/config_thu.yaml --ckpt <PATH_TO_CKPT>

---

License

This project is licensed under the Apache License 2.0, with the exception of certain components derived from or influenced by third-party projects such as Gaussian Splatting and VGGT.

• Apache License 2.0: All original code written for this repository is released under the Apache 2.0 license. See LICENSE.
• Gaussian Splatting: Some parts of this project are based on or adapted from Gaussian Splatting, which is licensed for non-commercial research use only. Please check the original project page and related license terms before use.
• VGGT: This repository also includes or depends on components related to VGGT. Please refer to the original VGGT repository for its license terms and usage conditions.

This repository is intended for academic research purposes. Users are responsible for checking and complying with the licenses of all included or referenced third-party components.

Citation

If you use this codebase, please cite our paper:

@InProceedings{Kim_2026_CVPR,
    author    = {Kim, Jingi and Kim, Wonjun},
    title     = {Generalizable Human Gaussian Splatting via Multi-view Semantic Consistency},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Findings},
    month     = {June},
    year      = {2026},
    pages     = {139-148}
}