PSGAN icon indicating copy to clipboard operation
PSGAN copied to clipboard

white spots in the generated image

Open melody-rain opened this issue 3 years ago • 4 comments

49_2048_fake

When training with the code of this repo, my generated images always have white spots. Does anyone have the similar results? And how to avoid it?

Thanks

melody-rain avatar Aug 30 '21 02:08 melody-rain

I meet similar problem in other GAN model. My generated images always have green spots at same location of face. I have no idea about the reason... Any suggestion wil be appreciated. Thanks.

PennyPeng369 avatar Sep 02 '21 06:09 PennyPeng369

It might be caused by instance normalization as hypothesized in https://openaccess.thecvf.com/content_CVPR_2020/papers/Karras_Analyzing_and_Improving_the_Image_Quality_of_StyleGAN_CVPR_2020_paper.pdf .

wanzysky avatar Sep 02 '21 09:09 wanzysky

@PennyPeng369 @wanzysky yes according to stylegan2, replacing all conv2d with demodulated conv2d, the white spots disappear in my experiments. But I am wandering why the author’s released model has no such a problem.

melody-rain avatar Sep 02 '21 09:09 melody-rain

49_2048_fake

When training with the code of this repo, my generated images always have white spots. Does anyone have the similar results? And how to avoid it?

Thanks

how many epo have you trained to get this sample? Does it happen to all the images?

AloneGu avatar Nov 03 '21 01:11 AloneGu

@PennyPeng369 @wanzysky yes according to stylegan2, replacing all conv2d with demodulated conv2d, the white spots disappear in my experiments. But I am wandering why the author’s released model has no such a problem.

Yes, the white spot is caused by instance normalization. Replacing the original conv2d could be a good idea.

wtjiang98 avatar Dec 14 '22 03:12 wtjiang98