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

why does G's input is (noise = opt.noise_amp*noise_+prev, prev)? #65

Closed yael-vinker closed 4 years ago

yael-vinker commented 4 years ago

Hi and thanks for sharing the code (:

According to the article: "The input to Gn is a random noise image zn, and the generated image from the previous scale x˜n, upsampled to the current resolution (except for the coarsest level which is purely generative)"

From the code (if I understood correctly): except of scale 0, all inputs to G are from the form: noise = opt.noiseamp*noise+prev G(noise, prev) (where prev is the output of the prevuois layer) when displaying G's input, it seems like "noise" is very similar to prev (which means that it is not a random noise image zn)

I would appreciate your explanation, what am I missing here? (: Thanks!