Random noise added to real and simulated patch before passing Discriminator

[a2bc]

Hi, Thanks so much for this paper and code. I track the code and see that: In kernelGAN.py d_pred_fake = self.D.forward((g_output + torch.randn_like(g_output) / 255.).detach())

In data.py

if not for_g:  # Add noise to the image for d
            crop_im += np.random.randn(*crop_im.shape) / 255.0

there is a Gaussian random noise with mean = 0 and std = 1/255 added to both the fake and real patches before passing to the Discriminator.

I read the paper and don't see any information about this trick. Can you give me some hints/motivations why you need this noise ? Thanks !

[sefibk]

of course - this is super important for GANs and images. Since real pixel values are integers, they can be ONLY very specific values between 0-1 (i.e. 1/255, 2/255, 3/255...) Therefore, a discriminator can distinguish between real and fake images easily by determining: if all pixels are n/255 exactly it is real, o.w. fake.

This is why you should add noise to "not allow" the discriminator have such an easy task

Is this clear?

[a2bc]

of course - this is super important for GANs and images. Since real pixel values are integers, they can be ONLY very specific values between 0-1 (i.e. 1/255, 2/255, 3/255...) Therefore, a discriminator can distinguish between real and fake images easily by determining: if all pixels are n/255 exactly it is real, o.w. fake.

This is why you should add noise to "not allow" the discriminator have such an easy task

Is this clear?

Totally make sense ;-) Thanks for the quick feedback !