SRGAN is a new program that attempts to do for photographs what Waifu2x does for drawings; upscaling photos while preserving sharpness and reconstructing lost detail. It only supports 4x upsacling, meaning that 16 new pixels are generated for every old pixel.
You have 2 options for playing with it yourself:
The hard way: If you have Linux and are willing to install Tensorflow, you can download an implementation of it here. It comes with a pre-trained network so you don’t have to do that yourself, though you can if you wish. https://github.com/brade31919/SRGAN-tensorflow
The easy way: use the website Let’s Enhance. https://letsenhance.io/ It also uses SRGAN, and the results are superior to SRGAN-tensorflow’s defaults. I’m unsure if this is due to modifications to the program itself, or it was trained on a better network. The first 20 images are free, afterwards it’s $6 a month or $55 a year.
Unfortunately, it’s a bit inconsistent. For every acceptable result, there’s several failure cases like these (tried with both LE and SRGAN-tensorflow). It may help to train the network on high res CGI images or game textures or or something, but I would likely need thousands of images and it would take a couple of weeks to train. Should I give it a shot anyways?
EDIT: Do not click on the “srgan” tag.
