What's the highest quality resizer in existence?

[Chainmax]

For Avisynth, excluding the superresolution functions, we have some pretty good solutions for ~250% resizing. I was wondering if a solution exists (commercial or not) that can resize to bigger factors while still being comparable to the original in terms of detail.

[foxyshadis]

I'd say above 200%, maybe 300% for some things, it stops being about "quality" and starts being more about "taste". The source has no more hidden detail to offer up, the tricks for manufacturing detail get exposed, and the only thing left to do is pour on grain or fake scribbles. There are also differences in whether noise is supressed, kept, or even enhanced; different images respond best to different noise treatments.

Lesser resizers get exposed as the super-lanczos with halos they are, spline/vector/fractal combos (photozoom, qimage, genuine fractals, blow up, pxl smartscale) start looking more like faceted gemstones, pure vector engines (SAR, Qimage Vector, EEDI2) have bad vectors made glaringly obvious.

If you're curious, the results of all the current commercial resizers are hosted here: http://www.americaswonderlands.com/digital_photo_interpolation.htm

You're welcome to choose the one that best matches your tastes. (Note that while most are reasonably slow, some are obnoxiously slow.) You could always try merging the output of a few different ones and see if that further improves things. :p

[Chainmax]

Thanks for the link, now I got me some eye-burning to do ;) :p.

[Waleska]

Extensis pxl SmartScale.

[Chainmax]

The person that did that comparison page either has no clue how to make a comparison or at the very least he didn't care. Not only are the enlargements JPGs when they should be PNGs, they are different sizes and don't contain the exact same area. What a shame.

[CruNcher]

I like the ones that clear the chroma stepping and sharpen a bit, over those that show chroma stepping and blur to much.
SmartScale and Photozoom Pro 2 :)

[*.mp4 guy]

Lad Deconvolution (http://www.general-cathexis.com/interpolation.html) is the highest quality resizer in existence, its also the slowest, so it isn't very usefull. The higest quality resizers that are usefull, are (imo) modified b-spline and Jensen-Zhao-Xin Li, with b-spline having a large speed advantage while still looking good and Jensen-Zhao-Xin Li having a large quality advantage while still being fast enough to use (computation time per frame measured in seconds, not hours or days).

[edit] for reference here (http://img170.imageshack.us/img170/3931/multiswarth7.png) is the best result I got from the current version of MultiSWAR I'm working on.

[Chainmax]

Wow, I remember that page. That's a well done comparison, thanks for the link :).
About LAD, I just googled and a thread in steve's digicam's forum mentioned that it could take weeks for it to process an image, why is that and where could it be useful then?

[foxyshadis]

Jensen-Zhao-Xin Li (and other variants of Xin Li) tends to look very empty at high zoom, and tends to have an overly high false-positive rate for edges. With cartoon-type stuff it's a nearly perfect resize, but grain and high frequency detail throws a total monkey wrench in it.

It's too bad that LAD Deconvolution requires so much guessing and retesting to get really good results. And that's just the parameters that make sense; a number of them seem to require intimate knowledge of the algorithm to even guess at. If you can get them right it offers the possibility of being perfect though, or as close as you can get with roundoff errors.

[Chainmax]

While we're at it, here's an old avatar I made out of a photo:

http://img57.imageshack.us/img57/7074/d9oaaj6.png (http://imageshack.us)

and here's a first attempt at resizing:

http://img206.imageshack.us/img206/4241/d9oav20qc4.png (http://imageshack.us)

Do you think I could get better results by trying LAD, EKG or MultiSWAR?

[akupenguin]

About LAD, I just googled and a thread in steve's digicam's forum mentioned that it could take weeks for it to process an image, why is that and where could it be useful then?

The weeks of computation time in that thread was to upscale a 8 MP photo to big enough for a 24"x36" print, i.e. somewhere on the order of 100-1000 MP. If you're not dealing with such ridiculously large images, the computation would be proportionately faster.

As fow why. Though I haven't ever used SAR Image Processor and afaict the author hasn't published the algorithm details, I infer the following from its website (http://www.general-cathexis.com/):
LAD deconvolution is similar to backprojection methods, in that it iteratively refines a big image to reduce the difference between the original and a downscaled copy of the big image. What LAD does differently (and slowly) is that each new candidate big image is itself iteratively generated by backprojection with a different metric.
See also Expectation Maximization (http://en.wikipedia.org/wiki/Expectation_Maximization) for a more general form of such algorithms.

[Chainmax]

That's very interesting, thanks for the explanation :).

[Terka]

is it worth to resize DV source 720x576 to 1440x1152
to "gain a quality" (resize before encoding)
instead resize while watching on my PC to 1280x1024
(my monitor default)
if its quality worth -which resizer do you suggest for good speed/quality

[foxyshadis]

No. If you had a huge wide 1920x1200 monitor, yes, it would make sense, and maybe even for 1650x1050 if it really mattered to you. But the upsize amount is too small to make a real difference in your case, if you use a bicubic renderer (VMR9 in mpc, or Haali's), and even weaker systems can handle spline/lanczos to 1280 in ffdshow pretty well. You'll spend an enormous amount of time and drive space for a very questionable result otherwise.