HD-Net

This repository contains the official implementation of "Face Inverse Rendering via Hierarchical Decoupling" PDF, which appeared in TIP 2022.

From top to bottom are the input face, predicted albedo, predicted shading, visualized light and predicted normal.

Training

Please see each subsection for training on different datasets. Available training datasets:

• 250k synthetic face images
• DPR
• FFHQ
• [Real Face Image] (Real paired face image needs warped and unwarped function for alignment from Nonlinear_Face_3DMM)

Generating the training and test csv

Run

python generate_dataset_csv.py \
   --dataset_path $DATA_DIR

Training

Run

python train_main.py \
   --dataset_path $DATA_DIR \
   --batch_size 8 \
   --log_dir  $LOG_DIR  \
   --stage $STAGE_NAME \
   --datadir $DTU_DIR

Testing on sample data

Run

python test_main.py 

Thanks to SfSNet-PyTorch, DPR, Nonlinear_Face_3DMM and 3DMM our code is partially borrowing from them.