PointSplat: Compact Gaussian Splatting via Human-Centric Prediction

ECCV 2026

Yujie Guo¹, Yudong Jin¹, Lingteng Qiu³, Zehong Shen¹, Zhen Xu¹, Jing Zhang², Xianchao Shen², Hujun Bao¹, Sida Peng¹, Xiaowei Zhou^1†

¹Zhejiang University    ²ByteDance    ³CUHK, Shenzhen

^†Corresponding author

⚙️ Installation

Create a conda environment and install the dependencies:

conda create -n pointsplat python=3.10 -y
conda activate pointsplat
pip install -r requirements.txt

📦 Checkpoint

Download the pretrained checkpoint with:

hf download Yujie0012/PointSplat_pretrained_weights pointsplat_mixed.pt \
  --local-dir pretrained

You can also download the checkpoint manually. Either way, the final layout should be:

pretrained/
└── pointsplat_mixed.pt

🚀 Inference

Inference on DNA-Rendering

Step 1: Download example data.

Download the DNA-Rendering example package from PointSplat_example_data with:

hf download Yujie0012/PointSplat_example_data DNA_Rendering_example.zip \
  --repo-type dataset \
  --local-dir datasets
unzip datasets/DNA_Rendering_example.zip -d datasets

Expected data layout:

datasets/DNA_Rendering_example/
├── validation_index.json
├── scenes/
│   └── <scene_id>/
│       ├── transforms.json
│       └── images/
│           └── <view_id>/
│               └── 000030.webp
└── masks/
    └── <scene_id>/
        └── fmasks/
            └── <view_id>/
                └── 000030.png

Tip

To test the model on more DNA-Rendering scenes, follow the processed DNA-Rendering dataset instructions in Diffuman4D.

Step 2: Run the inference script.

CUDA_VISIBLE_DEVICES=0 python -m src.main \
  +experiment=pointsplat_renbody \
  output_dir=experiments/eval_pointsplat_dna_rendering \
  dataset.roots='[datasets/DNA_Rendering_example]' \
  checkpointing.pretrained_encoder=pretrained/pointsplat_mixed.pt

Inference on THuman2.0

Step 1: Download example data.

Download the THuman2.0 example package from PointSplat_example_data with:

hf download Yujie0012/PointSplat_example_data THuman2_0_example.zip \
  --repo-type dataset \
  --local-dir datasets
unzip datasets/THuman2_0_example.zip -d datasets

Expected data layout:

datasets/THuman2_0_example/
└── val/
    ├── img/
    │   └── <scene_id>_<view_id>/
    │       └── 2.jpg
    ├── mask/
    │   └── <scene_id>_<view_id>/
    │       └── 2.png
    └── parm/
        └── <scene_id>_<view_id>/
            ├── 2_intrinsic.npy
            └── 2_extrinsic.npy

Tip

For the full THuman2.0 training/test data, refer to the dataset preparation instructions in GPS-Gaussian.

Step 2: Run the inference script.

CUDA_VISIBLE_DEVICES=0 python -m src.main \
  +experiment=pointsplat_thuman \
  output_dir=experiments/eval_pointsplat_thuman \
  dataset.roots='[datasets/THuman2_0_example]' \
  checkpointing.pretrained_encoder=pretrained/pointsplat_mixed.pt

📚 Citation

If you find this code useful for your research, please consider citing:

@article{guo2026pointsplat,
  title={PointSplat: Compact Gaussian Splatting via Human-Centric Prediction},
  author={Yujie Guo and Yudong Jin and Lingteng Qiu and Zehong Shen and Zhen Xu and Jing Zhang and Xianchao Shen and Hujun Bao and Sida Peng and Xiaowei Zhou},
  journal={arXiv preprint arXiv:2606.32036},
  year={2026}
}