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
3.3k stars 610 forks source link

Superresolution training and inference #121

Open surajp92 opened 4 years ago

surajp92 commented 4 years ago

I was looking at the use of SinGAN for the superresolution of images. Usually, when I am using architectures like SRGAN, there is a training and inference stage. But, for the SinGAN superresolution example, do I need to only give the low-resolution image? Do we not need to provide the high-resolution image for training the network? How is the network trained for SinGAN superresolution tasks?

Thank you.

tamarott commented 4 years ago

Only the LR image is needed for training. At inference time we continue to upsample the image using the trained model scale. Please see the paper for detailed explanation (section 4)
https://openaccess.thecvf.com/content_ICCV_2019/papers/Shaham_SinGAN_Learning_a_Generative_Model_From_a_Single_Natural_Image_ICCV_2019_paper.pdf