2013-06-16, 12:37
It is like "always driving 100 km/h" in the first gear :-)
There is basically no need to upscale movies that are already 1920 or near 1920 in width with lanczos 3 it is just iteratively calculating on the image with the nearlyidentical result you pushed in.To see how lanczos3 works see here: http://en.wikipedia.org/wiki/Lanczos_resampling Consider the sinc "evil" in domain space and see the polynoms that are used to calculated the convolution.
Now think of x \approx i in the multidimensionalInterpolation formula, you will realize that the result in order to be smooth "in subpixel dimension" as you your target area consists nearly only supporting points (i) + some minor per cent missing (x), that you will have approximately the same results when using bilinear, perhaps + some very small gaussian kernel + another very small edge enhancement.
In short: no need to calculate in O(n^2)*Convultion which is here a bit better than O(n^2) cause of the properties of the sinc.
In short: Set it to 20% and save the planet :-)
Edit:On could tune the "iterations" concerning the scaling width - but did not look into the code in detail to see if this is possible, e.g. stop after only one iteration - hehe, you see what you get if you do this?
There is basically no need to upscale movies that are already 1920 or near 1920 in width with lanczos 3 it is just iteratively calculating on the image with the nearlyidentical result you pushed in.To see how lanczos3 works see here: http://en.wikipedia.org/wiki/Lanczos_resampling Consider the sinc "evil" in domain space and see the polynoms that are used to calculated the convolution.
Now think of x \approx i in the multidimensionalInterpolation formula, you will realize that the result in order to be smooth "in subpixel dimension" as you your target area consists nearly only supporting points (i) + some minor per cent missing (x), that you will have approximately the same results when using bilinear, perhaps + some very small gaussian kernel + another very small edge enhancement.
In short: no need to calculate in O(n^2)*Convultion which is here a bit better than O(n^2) cause of the properties of the sinc.
In short: Set it to 20% and save the planet :-)
Edit:On could tune the "iterations" concerning the scaling width - but did not look into the code in detail to see if this is possible, e.g. stop after only one iteration - hehe, you see what you get if you do this?