best upscale algorithm

[luquinhas0021]

what is more accurate (retention of every detail): smart edge 2 or m spline (neural spline)? And why? Smart edge 2 creates details?

[LoRd_MuldeR]

Please edit your question in accordance with rule #12!
http://forum.doom9.org/forum-rules.htm

Also, you should try to be as specific as possible. What exactly are you trying to achieve? What is nature of your source material?

[luquinhas0021]

i`m talking about photo upscaling and single frame upscaling

[raffriff42]

Some algorithms might do a pretty nice job on some images, but none can be perfect, as the basic information is not present in the source. Therefore, there is no "best method." You must compare using your source material and make your own judgement.

Here's a good page for quickly comparing upscaling algorithms:
Video resize methods comparison (http://www.infognition.com/articles/video_resize_shootout.html) (infognition.com)
Click on the top-left thumbnail, named "smiths.avi", then the right popup.
Then click the buttons to compare the different algorithms.

Here's another one:
Comparisons of Image Magnification Methods (http://www.general-cathexis.com/interpolation/) (general-cathexis.com)
Hover your mouse over the numbers placed around the outside of the large image.

[pandy]

what is more accurate (retention of every detail): smart edge 2 or m spline (neural spline)? And why? Smart edge 2 creates details?

From math perspective: point resampling (Nearest-neighbor interpolation). All details are kept intact.

[Motenai Yoda]

From math perspective: point resampling (Nearest-neighbor interpolation). All details are kept intact.

Only if you are up-scaling by an exact multiple, ie 1.5x give you alternately 1 pixel wide column/right and a 2 pixel wide column/right.

anyway it adds steps that there aren't in source.

[feisty2]

padding/addborders
they blow the frame size up without doing anything to the actual clip :)

[pandy]

Only if you are up-scaling by an exact multiple, ie 1.5x give you alternately 1 pixel wide column/right and a 2 pixel wide column/right.

anyway it adds steps that there aren't in source.

Topic is about upscale not general resampling.
Yes fraction upscaling with point resampling can give described result but it is fully reversible i.e. non lossy.

[*.mp4 guy]

Point scaling is only non-lossy If you don't do /anything/ with the result. Beyond that, all of the major linear interpolators in avisynth use strictly interpolating kernels, thus they can be made to be lossless under any condition during which pointresize could be lossless by choosing the correct pixel offsets. For integer interpolation be odd factors, eg. 3, 5 you don't even need to calculate an offset, the center pixel ( ie. b, when scaling by a factor of 3 from the pixel set [1->a b c], is unchanged, b=1). This property is generally much much more useful, as lnear interpolation doesn't cause problems for subsequent operations the way pointscaling does. Depending upon how you upscale with nnedi3, it too can produce a lossless result, preserving all of the original picture information in a manner analogous to the situation with linear interpolation.

Point being, if you need "lossless" interpolation, you've got better options than pointresize.

[luquinhas0021]

*mp4 gui, tell me how i can make a lossless resize with nnedi3, once even spline 100 was better than it in images and videos that i had upscaled! Three more questions: antialiasing filter makes detail loss? Is there some detail lossless aa filter? What is better center shift correction algorithm that i can choose for nnedi3 parameter? nnedi3 doesn`t is update by tritical since 2011. It was discontinued?

[huhn]

if you change something by adding or removing something you lose something it's that simple.

[Gavino]

all of the major linear interpolators in avisynth use strictly interpolating kernels, thus they can be made to be lossless under any condition during which pointresize could be lossless by choosing the correct pixel offsets.
Yes.
The exceptions to this (ie those that are not strictly interpolating) are BicubicResize with b/=0 and GaussResize with values of p less than about 80.

See also this old thread: Lossless resizing.