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

Using SinGAN to generate hyper-spectral images (ch>3) #78

Closed FamousDirector closed 4 years ago

FamousDirector commented 4 years ago

Hello. I have modified the SinGAN architecture to work with hyper-spectral images (ch>3). When working with images with 5 channels, the synthetic samples generated were very realistic and worked well. However when I got up to 8 channel images, the generated samples began to degrade in quality. I've tried tweaking the network with the many exposed hyper-params available, but I hit other failure cases such as mode collapse.

I was wondering that the experts could provide some suggestions for improving my results with hyper-spectral images.

tamarott commented 4 years ago

Did you try to increase the generator and discriminator capacity? I guess this task is more challenging than generating RGB images.

FamousDirector commented 4 years ago

Increasing both the nfc and the number of layers in the discriminator/generator helped, but it took some tuning to prevent mode collapse. In the end the best solution I had was actually changing the representation of my multi spectral images. Thanks!