search

tamarott / SinGAN

Official pytorch implementation of the paper: "SinGAN: Learning a Generative Model from a Single Natural Image"
https://tamarott.github.io/SinGAN.htm
Other
3.31k stars 611 forks source link

Question about paper #71

Closed jianwang-scu closed 4 years ago

jianwang-scu commented 4 years ago

In Training section:

The reconstructed image has another role during training, which is to determine the standard deviation σn of the noise z_n in each scale

What that suppose to mean? BTW, thanks for your excellent work! 💯

tamarott commented 4 years ago

We want the additive noise at each level to be proportional to the amount of missing details within the image. The RMSE between the upsampled reconstructed image from the previous scale and the real image can provide such an assessment. So the STD of the additive noise at each of SinGAN's level is proportional to this.

jianwang-scu commented 4 years ago

thanks!