[Lecture 12] GAN generating fourier domain image #103

In FFT, I understand frequency(fourier) image is another representation of original image, also reconstruct original image. I think model with FFT can't do end-to-end leaning. But frequency image contains geomatrical properties of image and convolution operation can be conducted in there. So I think it is enough to fit model with frequency image. Is it possible to make auto encoder(or GAN) to generate frequncy domain image?

Maybe without measuring it is plausible to human vision, same problem of generating face from auto encoder.