QuerySplat: Decoupling Geometry and Appearance Representations in 3DGS Prediction
Project Page | Paper | GitHub
QuerySplat is a feed-forward 3D Gaussian Splatting (3DGS) framework driven by geometric priors and explicit appearance decoupling. Specifically, it uses a dual-branch query-based decoder where the geometry branch leverages a pretrained Vision Geometric Model for spatial understanding, while the appearance branch recovers high-frequency details through a dedicated pathway separated from geometric attribute regression.
Installation
git clone https://github.com/inspatio/QuerySplat.git
cd QuerySplat
conda create -n querysplat python=3.12 -y
conda activate querysplat
# Tested configuration: PyTorch 2.11.0 + CUDA 12.8.
python -m pip install torch==2.11.0 torchvision==0.26.0 \
--index-url https://download.pytorch.org/whl/cu128
python -m pip install --no-build-isolation -r requirements.txt
python -m pip install -U huggingface_hub
Inference
Place any number of images from one scene in --input_folder. Run inference with Test-Time Optimization (TTO):
python -m scripts.infer \
--config checkpoints/querysplat_vggto_1B_512_8192.yaml \
--checkpoint checkpoints/querysplat_vggto_1B_512_8192.safetensors \
--input_folder data/my_scene \
--output_dir outputs/my_scene \
--use_tto
Omit --use_tto to run the feed-forward model without test-time optimization.
Citation
@article{li2026querysplat,
title={QuerySplat: Decoupling Geometry and Appearance Representations in 3DGS Prediction},
author={Li, Yinglong and Shen, Donghui and Zhang, Xiaoyu and Ye, Zhichao and Wu, Hongyu and Hao, Aimin and Zhang, Guofeng and Liu, Haomin},
journal={arXiv preprint arXiv:2608.01186},
year={2026}
}