VideoSR, which trained for YCbCr sequence, to handle YCbCr video without any conversion.
However, currently version only support Y component.
The channel of Y, Cb and Cr is all 1, therefore you can test it by using Cb and Cr, but it's result is optimised to Y component.
About VideoSR
VideoSR is a part of the "Sharpic", which automatically generate HQ sequences from LQ sequences. (Less than 2K).
NVIDIA GPU & CUDA
pytorch 1.12
Time consumption (NVIDIA Tesla P100): (24+ Hours)
- YUVESRGAN (BasicSR)
Test dataset: Flickr2K 2451-2650
| Metrics | Anchor(Original RealSR) | Real-YUV |
|---|---|---|
| PSNR-Y | 25.595302 | 25.845682 |
| MSSIM |