Learning Super-Resolution space

The goal of this work is learning the super-resolution space, which is one of the challenges presented in NTIRE 2021 competition. The mission was to solve the challenge according to the competition rules - given low resolution image, to produce good quality super-resolution image.
In reality, many high-resolution images can be downsampled to the same low-resolution image.
The challenge focused on producing arbitrary number of super-resolution images capturing meaningful diversity, using the same input of low-resolution image. in addition, the output images need to be consistent to the input, and with high photo-realism as perceived by humans.
We created a model, based on SR-GAN, which can get as an input both low-resolution image and random noise, and produce arbitrary number of super-resolution outputs with meaningful diversity, and consistent to the original low-resolution image.

space of plausible SR

Learning Super-Resolution space

Collaboration:

Challenge of NTIRE 2021