Code for the paper Global Texture Enhancement for Fake Face Detection in the Wild, CVPR 2020.
Authors: Zhengzhe Liu, Xiaojuan Qi, Philip H.S. Torr
Download the model and the demo data for StyleGAN-FFHQ, StyleGAN-CelebA and PGGAN-CelebA, respectively. In each folder,
python demo.py
It will print the file name, processing (resize 8x, JPEG or original image) and prediction (0 for fake and 1 for real).
Download the 10k images from FFHQ, CelebA, and generate 10k images using StyleGAN, PGGAN on the datasets. Please save the 1024*1024 resolution images with PNG format, not JPG.
Optionally, to evaluate the low-resolution GANs, download images from CelebA dataset, and generate images using DCGAN, DRAGAN and StarGAN.
Generate the filelist for training.
python gene.py
Put graminit.pth to the training folder as initialization, and start training, while evaluate the model on the validation set regularly to choose the best model.
python main.py
Modify "root" folder and image path in test.py, and then test the images on all the datasets.
python test.py
python test2.py
python test3.py
If you have any questions or suggestions about this repo, please feel free to contact me (liuzhengzhelzz@gmail.com).