This is part of the Deep Portrait Relighting project. If you find this project useful, please cite the paper:

  title={Deep Single Portrait Image Relighting},
  author = {Hao Zhou and Sunil Hadap and Kalyan Sunkavalli and David W. Jacobs},
  booktitle={International Conference on Computer Vision (ICCV)},


This code is not optimized and may not be well organized.


3DDFA: https://github.com/cleardusk/3DDFA (download the code and put it in useful_code, follow the instruction to download model and setup the code)

Environment setup:

I use miniconda to setup virtual environment

  • Create a virtual enviroment named RI_render (you can choose your own name): conda create -n RI_render python=3.6
  • Install pytorch: conda install pytorch torchvision cudatoolkit=9.2 -c pytorch -n RI_render
  • Install dlib: conda install -c conda-forge dlib -n RI_render
  • Install opencv: conda install -n RI_render -c conda-forge opencv
  • Install scipy: conda install -n RI_render -c conda-forge scipy
  • Install matplotlib: conda install -n RI_render -c conda-forge matplotlib
  • Install cython: conda install -n RI_render -c anaconda cython
  • Compile 3DDFA as mentioned in the github webpage
  • Compile cython in utils/cython, follow the readme file
  • Install Delaunay Triangulation:
  • Install libigl:
  • Install shtools: https://github.com/SHTOOLS/SHTOOLS
  • Install cvxpy: conda install -c conda-forge cvxpy

Steps for rendering

  1. fitting 3DDFA: run bash run_fit.sh, will generate several files in result:
    *_3DDFA.png: draw 2D landmark on face
    *_depth.png: depth image
    *_detected.txt: detected 2D landmark on faces
    *_project.txt: projected 3D landmark
    *.obj: fitted mesh

  2. run bash run_render.sh
    generate albedo, normal, uv map and semantic segmentation:
    *_new.obj: obj file for rendering
    in render:
    *.png show generate images
    *.npy show original file of albedo, normal, uv map and semantic segmentation. NOTE: if you can install OpenEXR, you can save npy as .exr file

  3. run bash run_node.sh
    Apply arap to further align faces
    in render:
    generate arap.obj an object of arap algorithm
    *.node and *.ele temperal files for applying arap

  4. run bash run_warp.sh
    create warped albedo, normal, semantic segmentation in result/warp:

  5. run bash run_fillHoles.sh
    remove ear and neck region and fill in holes in generated normal map:
    create full_normal_faceRegion_faceBoundary_extend.npy and full_normal_faceRegion_faceBoundary_extend.png in result/warp

  6. run bash run_relight.sh
    relighting faces
    download our processed bip2017 lighting through (https://drive.google.com/open?id=1l0SiR10jBqACiOeAvsXSXAufUtZ-VhxC), change line 155 in script_relighting.py to poit to the lighting folder
    Apply face semantic segmentation to get skin region of the face: https://github.com/Liusifei/Face_Parsing_2016 save the results in folder face_parsing/ (examples are shown in face_parsing, you can also skip this by adapting the code of script_relighting.py)