Blender rendering script for FaceScape dataset accompanying paper "NeLF: Neural Light-transport Field for Portrait View Synthesis and Relighting"

• Blender

• PyTorch & Kornia - for downsampling

• numpy

• OpenCV-Python

• matplotlib

The following steps denote how to generate training data for NeLF.

0. Prepare dataset. Download TU-Model from FaceScape dataset. Extract them to a desired folder. Example data structure:

   [path_to_model_folder] --- 1 --- dpmap
                           |     |- models_reg
                           |- 2
                           |- 3
                           ...
   

1. Remove the red cap texture in FaceScape dataset by running:

   python remove_hat.py --model_folder [path_to_model_folder] --out_folder [path_to_output_model_folder]
   

2. Prepare environment map. There are some free environment maps available online. For example, the Laval Indoor HDR Dataset. Download and extract them to a desired location.

   Example data structure:

   [path_to_envmap_folder] --- 0.hdr
                            |- 1.hdr
                            ...
   

3. (Optional) Sometimes the environment maps are too dimmed. They can be adjusted with a normalization script:

   python normalize_env.py --folder [path_to_envmap_folder] --out_dir [path_to_output_envmap_folder]
   

4. Randomly rotate the environment map for self rotation training

   python random_rotate.py --root_dir [path_to_envmap_folder] --out_dir [path_to_rotated_envmap_folder]
   

5. Set up config files. Change the paths in configs/default.txt to your corresponding folders.

6. Render models with Blender

   python batch_render_default.py --config configs/default.txt
   

7. Prepare downscaled environment map for the outputs

   python prepare_env_map.py --root_dir [root_folder_of_output]
   

8. Collect dataset

   python collect_data_default.py --root_dir [root_folder_of_output] \
   --out_dir [output_folder]
   

(Optional) Generate additional data for IBRNet and SIPR

1. Set up config files. Change the paths in configs/relight.txt to your corresponding folders for SIPR relighting training data and in configs/uniform.txt for IBRNet view synthesis training data.

2. Render models with Blender

   For relighting:

   python batch_render_relight.py --config configs/relight.txt
   

   For view synthesis:

   python batch_render_uniform.py --config configs/uniform.txt
   

3. Prepare downscaled environment map for the outputs

   python prepare_env_map.py --root_dir [root_folder_of_output]
   

4. Collect dataset

   For relighting:

   python collect_data_relight.py --root_dir [root_folder_of_output] \
   --out_dir [output_folder]
   

   For view synthesis:

   python collect_data_uniform.py --root_dir [root_folder_of_output] \
   --out_dir [output_folder]