The models presented in the paper are trained on class-conditional ImageNet (where the input is Gaussian noise and one of 1000 classes, e.g., "car") and unconditional FFHQ (where the input is only Gaussian noise).