FSRCNN, extended with local residual learning and recursive blocks.
All networks have been trained with --distort=True. Means they can reduce slight compression artifacts. There is a version of this network trained to remove strong compression artifacts here.
LineArt version has been trained using a dataset that only consists of flat lines, so it may look like an oil painting on a more natural type of content. Need a better dataset.
If you placed this shader in the same folder as your mpv.conf, then your config should look like this (regardless of the operating system):
profile=gpu-hq
glsl-shader="~~/FSRCNNX_x2_8-0-4-1.glsl"
For those evaluating Super Resolution quality, I recommend SSIM with Mean Absolute Deviation (MAD) pooling (rather than default mean pooling) or IW-SSIM. For a more modern approach, PieAPP is one of the best CNN-based metric for capturing human preference in Super Resolution. I do not recommend using PSNR or MS-SSIM (including DSSIM and SSIMULACRA2), as these metrics are much less effective for Super Resolution tasks.