[Alexkral]

I was referring to the metrics (psnr and ssim), I didn't test the speed.

The VDSR network learns to estimate a residual image, which is the difference between a high-resolution reference image and a low-resolution image that has been upscaled to match the size of the reference image. The low-resolution image in usually upscaled using bicubic interpolation, but I have obtained better results using Lanczos. I trained several models with different scaling factors, which are used depending on the upscaling ratio. VDSR allows for multiscale training, but it hasn't worked well for me. So it is arbitrary ratio.

In AviSynth Shader, when using the Width and Height parameters to resize the texture, the c0 and c1 registers are set to the output values, so I need to use another parameter to set another register with the input values.