The virtual Waifu pictures are generate by AI using NVIDIA famous style GAN2 algorithm. The training set is composed of 2500 images generated by TWDNE website. Resolution of each image is 512 x 512.
Style GAN2 paper: https://arxiv.org/abs/1912.04958
TWDNE website: https://www.thiswaifudoesnotexist.net/
Style GAN2 for HD fake human face: https://github.com/RyanWu2233/Style_GAN2_FFHQ
NVIDIA official source code: https://github.com/NVlabs/stylegan2
Execute or include file named SG2_main.pu. Then execute following instruction:
model = GAN()Create style GAN2 object
model.train(restart=False)Train modelmodel.predict(24)Generate fake image with 4 rows and 6 cols
model.predict(6)Generate fake image with 2 rows and 3 cols
model.predict(2)Generate fake image with 1 rows and 2 cols
model.save_weights()Save model
mdoel.load_weights()Load model