[arXiv Paper]

This is the official code for FG 2024 paper.

WSDF learns a 3D Face Model from registered meshes through disentangling identity and expression without expression labels.

• 2024/04/15: initial commit

• clone

  git clone https://github.com/liguohao96/WSDF.git --recursive

• environment

  instal python packages through pip install -r requirements.txt. or manually install torch, numpy, opencv-python, pillow, imageio, plyfile, nvdiffrast

• face model

  FLAME

  please first down FLAME 2020 and FLAME texture space from here and put extracted files under Data/FLAME2020, the directory should be like:

  Data/FLAME2020
  ├── female_model.pkl
  ├── FLAME_texture.npz
  ├── generic_model.pkl
  ├── male_model.pkl
  ├── tex_mean.png 
  └── ...
  

• datasets

  CoMA please download Registered Data from original website and unpack into Data/Datasets/CoMA/COMA_data

  the directory should be like:

  Data/Datasets
  ├── CoMA
  │   ├── COMA_data
  │   │   ├── FaceTalk_xxx
  │   │   └── ...
  

• build cache for SpiralConv

  First, make sure 3rdparty/spiralnet_plus is downloaded (if not please run git submodules sync --recursive). Then, run

  python preprocessing/make_spiral_cache.py --name FLAME2020

  this will build sampling matrix and indices for SpiralConv and save at Data/SpiralCache_FLAME2020.pth.

• split datasets into train/valid/test

  CoMA

  bash preprocessing/coma/run.sh [file path to COMA_data]

  this will generate label files in Data/dataset_split/CoMA.

Train VAEs with specified config:

python train.py [config file] --batch_size 8

It will save intermediate result and final checkpoints at ./temp/[name]

• using trained checkpoint

  assuming checkpoints are downloaded at Data/checkpoints/coma_ep100.zip

  from utils.load import load_wsdf
  model_dict = load_wsdf("Data/checkpoints/coma_ep100.zip")
  
  print(model_dict.keys())
  # dict_keys(['Encoder', 'VAEHead', 'Decoder', 'NeuBank', 'Norm', 'Topology'])
  
  print(model_dict["Topology"].keys())
  # dict_keys(['ver', 'tri', 'uv', 'uv_tri'])
  
  norm, decoder  = model_dict["Norm"].cuda(), model_dict["Decoder"].cuda()
  i_code, e_code = torch.randn(1, decoder.dim_I, device="cuda"), torch.randn(1, decoder.dim_E, device="cuda")
  ver = norm.invert(decoder(i_code, e_code))  # 1, NV, 3
  
  # export
  from utils.io import mesh_io
  mesh_io.save_mesh("out.ply", (ver.squeeze(0).detach().cpu().numpy(), None, None, model_dict["Topology"]["tri"].cpu().numpy(), None, None))

First, make sure tiny-cuda-nn pytorch extension is installed through pip (pip install git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch) or manually following offical guide.

Then, generate initial texture through python preprocessing/make_tex_mean.py --name FLAME2020

Finally, generate 3D mesh and texture through

python applications/text_3d.py --checkpoint [checkpoint file] --save_gif True --save_ply True --config [config file] 

or

python applications/text_3d.py --checkpoint [checkpoint file] --save_gif True --save_ply True \
--text "a DSLR photo of [], ..." --negative_text "blur, ..." --tex_file Data/FLAME2020/tex_mean.png --name [name]

It will save intermediate result and final mesh at ./temp/APP-tex_3d/[name]

First, follow the first two steps in Text-to-3D (install tiny-cuda-nn and generate initial texture).

Then, download FLAME Mediapipe Landmark Embedding from here and put extracted files under Data/FLAME2020.

Finally, fitting image through

python applications/fit_2d.py --checkpoint [checkpoint file] --save_gif True --save_ply True --config [config file] 

or

python applications/fit_2d.py --checkpoint [checkpoint file] --save_gif True --save_ply True \
--fitting_image [image file] --tex_mean Data/FLAME2020/tex_mean.png --name [name]

It will save intermediate result and final mesh at ./temp/APP-fid_2d/[name]

Our network and evaluation codes are based on spiralnet_plus and FaceExpDisentanglement, thanks for their great works.

Our text-to-3d codes are based on Fantasia3D and Perp-Neg, thantks for their great works.

@misc{li2024wsdf,
  title  = {3D Face Modeling via Weakly-supervised Disentanglement Network joint Identity-consistency Prior}, 
  author = {Guohao Li, Hongyu Yang, Di Huang and Yunhong Wang},
  year   = {2024},
  eprint = {2404.16536},
}