Originally Posted by stevenjw
chroma upscaling: NNEDI3 32 neurons
image doubling: use NNEDI3 to double Luma resolution - only is scaling factor is 1.5x (or bigger)
image upscaling: Jinc - 3 taps + activate anti-ringing filter
image downscaling: Lanczos - 4 taps + activate anti-ringing filter
There's no "correct" or "incorrect". But I can give some comments. E.g. using NNEDI3 does help, with both image doubling and chroma upscaling. But the benefit of using NNEDI3 for image upscaling is higher than that for chroma upscaling. So if you have enough GPU power to spare, you can use NNEDI3 for both chroma upscaling + image doubling. However, it's worth considering whether increasing the number of NNEDI3 neurons for image doubling wouldn't bring more visible benefit than using NNEDI3 for chroma upscaling. I'm not sure what the correct answer is there, it might even depend on the content you're watching. Increasing the number of neurons has a small positive benefit, but it can sometimes be *very* small, so it's also possible that using NNEDI3 for chroma upscaling helps more than increasing the number of neurons for image doubling. In any case, if your GPU can do that, I'd at least use 32 neurons for image doubling. If your GPU can't do that, and NNEDI3 for chroma upscaling at the same time, then I'd rather drop NNEDI3 for chroma upscaling instead of lowering the neuron count for image doubling. 16 neurons is "ok", but 32 is often better. The quality difference between 64 and 32 is smaller than that between 16 and 32.
When using NNEDI3 image doubling, using Jinc for upscaling might not be necessary. A cheaper algorithm might do. But of course using Jinc won't harm, if your GPU has the power. For downscaling Lanczos4 + AR is fine. Alternatively, for large downscaling factors, you might want to consider using linear light (but only for downscaling, not for upscaling). Linear light downscaling produces stronger halos, though, because of which I'd use a 2-tap algorithm then, like Bicubic or Catmull-Rom, instead of Lanczos.