DveloperY0115 / TexGAN

Learning Texture Generators for 3D Shape Collections from Internet Photo Sets

TexGANPrivate

Source code of the paper "Learning Texture Generators for 3D Shape Collections from Internet Photo Sets" by Rui Yu, Yue Dong, Pieter Peers and Xin Tong from BMVC 2021.

Dependencies

tensorflow-gpu==1.14
opencv-python
tqdm

Inference

• Copy exp_config\config_{car,ffhq,shoe}.py to exp_config\config.py
• The entry of inference process is test_process/run_inference.py.
  Run this command:
  python test_process/run_inference.py --output_dir Path/To/Output/Folder --data_root Path/To/DemoInputData --model_path Path/To/PretrainedModelFile --dataset {car,ffhq,shoe}
  The results will be found in Path/To/Output/Folder/test_result

Training

• Copy exp_config\config_{car,ffhq}.py to exp_config\config.py
• Run this command:
  python train_process/train_chart6view.py --output_dir Path/To/Output/Folder --data_root Path/To/TrainingData --phase_kimg 200 --gpu_id 0,1,2,3 --batch_group 8 --syn_batch 1 --real_batch 1 --exp_type {car,ffhq} --grad_repeat 4

Cuhk Dataset

Pretrained model: PretrainedModel
DemoInputData: CarDemo
Training Data: CarCuhk_RealExp
Source shapes of DemoInputData: CarDemoSourceShape

FFHQ Dataset

Pretrained model: PretrainedModel
DemoInputData: FFHQDemo
Training Data: FFHQ_RealExp

Shoe Dataset

Pretrained model: PretrainedModel
DemoInputData: ShoeDemo
Due to copyright issue, we cannot release the training data of the Shoe dataset.