I'm curious about why the custom anime4k CNN is not using YUV input images, specifically focusing on the Y component. This approach is quite common in super resolution models due to the reduced sensitivity of our eyes towards chroma components.
The upscaling could be done only on the Y component and for the U/V components a simple bicubic upscale could be used, thus decreasing the computation complexity of the problem.
I'm curious about why the custom anime4k CNN is not using YUV input images, specifically focusing on the Y component. This approach is quite common in super resolution models due to the reduced sensitivity of our eyes towards chroma components.
The upscaling could be done only on the Y component and for the U/V components a simple bicubic upscale could be used, thus decreasing the computation complexity of the problem.