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

Input for different training iterations #60

Open xuwh15 opened 4 years ago

xuwh15 commented 4 years ago

Hello I have a little doubt about the paper. Since the training set only consist of one image, I believe the same image is downsampled and fed into the discriminator over and over again in different iterations? So is there any special tricks to make the network learn patch distribution from the same input over different iterations?

tamarott commented 4 years ago

In all the scales we use a patch-discriminator which has a small receptive field of 11x11, therefore the discriminator never sees the full image, but only 11x11 patches from it, and return a discrimination score for each patch. This means that practically, our training set is the set of all overlapping patches within the real image across multiple scales.