This is the official implementation of HGG: Learning Efficient and Generalizable Human Representation with Human Gaussian Model.

[2025.7] We released the codebase for HGG! Training code and inference code are available.

pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 --index-url https://download.pytorch.org/whl/cu118
pip install triton=2.1.0
pip install -U xformers --index-url https://download.pytorch.org/whl/cu118

git clone --recursive https://github.com/ashawkey/diff-gaussian-rasterization
pip install ./diff-gaussian-rasterization

pip install git+https://github.com/NVlabs/nvdiffrast

# LGM requirements
pip install -r requirements.txt 

# more dependencies
pip install numpy==1.26.4
pip install torch_cluster
pip install smplx
pip install git+https://github.com/facebookresearch/pytorch3d.git@stable

Our training dataset can be downloaded from MvHumanNet Repo

Please link the downloaded dataset to ./mvhuman_24, you can run the following command to prepare the dataset:

Please download the NEUTRAL pkl from SMPL model and put it as ./smpl_renderer/mvhuman_tools/visual_smpl/smpl_/SMPL_NEUTRAL.pkl .

Then prepare the connectivity of SMPL mesh:

python vertex_connectivity.py # first get the connectivity of the mesh, you will get a pt file

First, copy all the data that you want to test into ./test_dataset, like ./test_dataset/200151

Then, run the following command to test the model.

bash eval_mvhuman.sh 

You will get a workspace with rendered images.

First please list all the data, cameras and poses you want to train into ./mvhuman_24/human.json like:

{
    "human_id": [ "100001", "100002", "etc"],
    "view_id_list": [ "22236222",  "22236234",  "22236236", "etc" ],
    "pose_id_list": [ "0005", "0055", "0105", "etc"]
        
}

Then you can run the training script:

bash train.sh # modify the gpu1/8.yaml

You might also want to modify the gpu1/8.yaml file to set the training parameters to use multi-gpu training.

This work is generously supported by Tencent.

This work is built on the amazing work of LGM and MvHumanNet. We thank them for sharing the wonderful code and dataset!

If you find this code useful in your work, please consider citing our paper (and maybe give a star to this repo!):

@inproceedings{liu2025learning,
  title={Learning Efficient and Generalizable Human Representation with Human Gaussian Model},
  author={Liu, Yifan and Zhang, Shengjun and Dai, Chensheng and Chen, Yang and Liu, Hao and Li, Chen and Duan, Yueqi},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={11797--11806},
  year={2025}
}

1. When building the diff-gaussian-rasterization library, you may encounter the following error:

   The detected CUDA version (12.5) mismatches the version that was used to compile
      PyTorch (11.8). Please make sure to use the same CUDA versions.
   

   You might want to use this command to get to the correct CUDA version that we installed in previous steps:

   export CUDA_HOME=/usr/local/cuda-11.8

2. CUDA out of memory, trying to allocate 133792483MiB... in diff-gaussian-rasterization

   This is a known issue in the diff-gaussian-rasterization library. Please refer to the github issue