Resize 1080p to 720p?

[TheRyuu]

... use an avisynth script and spline36resize(1280,720) the video so you now have 720p...

As long as he just need to scale down less than half the original size, isn't it better spline16resize?

I think 8 samples too much for scale down only to 2/3 of full size, it maybe using spline36resize will lead to less sharpness than spline16resize. But I think hardly someone can notice the difference viewing the resulting video: the two videos needs be compared frame by frame

This makes me to think run some test.... never scaled 1080 to 720 (scaling 1080 to 320, more often, and spline36resize works like a charm )

Even though spline36 is a "sharp" re-sizer, it won't cause artifacts when downsizing. If anything spline36 will have a better "interpreted" image than spline16. Spline is better than Lanczos at downsizing (lanczos will causing ringing, blackman re-size (in avisynth) attempts to correct this). I've never heard of anything like using spline16 when you only need to scale down to less than half the original size which makes me think that's some made up thing.

Bicubic is also acceptable when downsizing (more speed). Side effect of downsizing with bicubic is some AA. It probably won't be as sharp as spline36.

[Mister Hatt]

Also, regardless of how good you are at math, if the function's programming itself is broken (spline16 and spline64) then it will NOT do a better job. This has been fixed in avs2.6 but I don't think most people here are using that yet.

Re math, yes, I also know how it works, and while not a math student, I have written my own resizers and fully understand the sampling systems used. When you are downscaling, you do not want to blur it, or you lose more data regardless. Hence why more iterations are appropriate. Upscaling however you will want more accurate sampling, and so going higher is advised. Not that upscaling is ever advised.

Additionally, seems you really DON'T understand the numbers in these filters. The tap number is for iterations, not samples. Spline16 is 4 taps, spline36 is 6, spline64 is 8, etc. IIRC they all use the same sample size and then depending on number of taps, shift the sample points for the next iteration. That is how they build the number of samples up. I don't recall why the tapcount is squared for the names though.*

*might in fact be wrong as I haven't read up on this in a while, but AFAIK that is the reason for it.

[fuzzypuff258]

Good info to know, thanks!