Project Page | Paper | arXiv
The main contributions of our human image systhesis model are listed as followed:
- The first method to synthesize semantic disentangled full-body human image.
- We propose a 3D-aware super-resolution module which enable our project achieve 3D-aware image synthesis at
$1024^2$ resolution.
- Training code
- Datasets prepocessing code
- Code for more applications
- We finished our training and testing tasks on 4 NVIDIA A40 GPUs.
- 64-bit Python 3.8 and PyTorch 1.11.0 (or later). See https://pytorch.org for PyTorch install instructions.
- CUDA toolkit 11.3 or later.
- pytorch3d. See https://github.com/facebookresearch/pytorch3d/blob/main/INSTALL.md for install instructions.
Clone this repo:
git clone https://github.com/Alfonsoever/SemanticHuman-HD.git
cd SemanticHuman-HD
We suggest use anaconda to manage the python environment:
conda env create -f environment.yml
conda activate HD
python setup.py install
sh ./scripts/download_model.sh
Download SMPL models (1.0.0 for Python 2.7 (10 shape PCs)) and move them to the corresponding locations:
mkdir training/deformers/smplx/SMPLX
mv /path/to/smpl/models/basicModel_f_lbs_10_207_0_v1.0.0.pkl training/deformers/smplx/SMPLX/SMPL_NEUTRAL.pkl
sh ./scripts/test.sh
sh ./scripts/fid.sh
![]() |
![]() |
![]() |
![]() |
![]() |
![]() |
To see more details,please go to our project page.