search

chuangchuangtan / NPR-DeepfakeDetection

94 stars 2 forks source link

question about 4-classes training setting #6

Closed YinqiCai closed 2 months ago

YinqiCai commented 2 months ago

Hi, , thanks for your great work! But I have a confusing question: What's the meaning of 4-classes training setting mentioned in the paper? I found it seems to be 2-classes(real/fake) training setting in the code.

chuangchuangtan commented 2 months ago

The term "4-classes training setting" refers to the amount of data in the training set. The training set of CNNDetection contains 20 classes of images, with fake images generated by ProGAN and real images sampled from the LSUN dataset. The Official ProGAN code release the pretrained models on LSUN dataset.

YinqiCai commented 2 months ago

I get it, thank you very much!