Commercial consumer Super-Resolution vs Avisynth solutions

[CruNcher]

Lets see and show where we are @ :)

Source = http://mirror05.x264.nl/CruNcher/subjective_evaluation/SR/source.avi
http://s1.directupload.net/images/100422/a8jzme2n.png

Target = 640x360 2x input

MotionDSP (GPU)= http://mirror05.x264.nl/CruNcher/subjective_evaluation/SR/MotionDSP-PC-reversed.mkv (H.264)
http://s5.directupload.net/images/100422/fxq7gu5q.png
(uncompressed result)

Nero (CPU)= http://mirror05.x264.nl/CruNcher/subjective_evaluation/SR/Nero.mkv (H.264)
http://s5.directupload.net/images/100422/76ulltbx.png
(uncompressed result)

Infognition (CPU)= http://mirror05.x264.nl/CruNcher/subjective_evaluation/SR/Infognition.mkv (H.264)
http://s10.directupload.net/images/100422/7wy43z82.png
(uncompressed result)

Avisynth (CPU) (CPU+GPU) = (post your results)

PS: Yes i know that the colors of the MotionDSP frame are wrong im currently trying to fix that.

[julius666]

spline36resize
http://s10.directupload.net/images/100422/jie4v4m3.png

nnedi2
http://s5.directupload.net/images/100422/baaf7on6.png

nnedi3
http://s10.directupload.net/images/100422/b77e3yiv.png

[Keiyakusha]

Does additional processing like sharpening allowed? Because as for me, nnedi beats Infognition however Infognition is sharper. So nnedi2/3+lsfMOD will do the job.

[CruNcher]

Yep nnedi beats it in terms of aliasing prevention (though in those part MotionDSP is also not bad @ all) and yes additional filter results are also allowed that's where the power of Avisynth comes from, as long as you mark it as a combined solution i see no problem :)

[julius666]

in terms of aliasing prevention (though in those part MotionDSP is also not bad @ all)

Well, MotionDSP looks like a blurred spline36resize to me :devil:
http://s5.directupload.net/images/100422/fxq7gu5q.pnghttp://s10.directupload.net/images/100423/seemxf48.png

spline36resize(640,360)
blur(1)

And Infognition is the same, just with a light sharpening:
http://s10.directupload.net/images/100422/7wy43z82.pnghttp://s5.directupload.net/images/100423/x7aj8y5a.png

spline36resize(640,360)
sharpen(0.25)

Nero's solution is somewhere between the two.
It's a bit funny, that they sell them as "super-resolution". :)

compare them with this:
http://s5.directupload.net/images/100423/zsr8w6am.png

the avisynth code:
nnedi3_rpow2(rfactor=2)
SSSharp(ssw=true, iter=3, strength=0.7, rad=0.75)

the difference between the one with ssharp and the one withouth:
http://s5.directupload.net/images/100423/zsr8w6am.pnghttp://s10.directupload.net/images/100422/b77e3yiv.png

[CruNcher]

You don't have to forget what Super-Resolution is it's a algorithm to recover lost detail by using multiple frame data http://en.wikipedia.org/wiki/Super-resolution :) (these recovered detail also construct the sharpness level)
Only showing screens unfortunately is only half the truth, that's why i provided some example H.264 encodes with the results of each, all encoded CRF so you can clearly see how complex the uncompressed input ended up in terms of bitrate and preserved details Infognition is amazing in that part, sure the main problem on the viewing side stays persistent (aliasing) that way ;)
Also you can clearly see that MotionDSPs aliasing is better then any of the current posted results, yes even needi3 (for that it artificializes the source completely also that way it's gaining the highest compression result) :)
Im pretty sure that Sharpened it would look very clean also in motion and non of the aliasing would be visible (though Vreveals Filter Chain is bad for such advanced stuff), the price is high for that result as you can see all the real data is lost (something that Super-Resolution shouldn't do, though MotionDSP gives you no controls over anything here) and you can see clearly temporal artifacts in the MotionDSP result look @ Brads finger, btw i still didn't figured out what's up with the colors in Vreveal but the output is coming like that from the GPU :)

The aliasing prevention isn't clearly seen from a picture it's absolutely needed to watch this in motion to rate it's efficiency and effect it has visually

And we yet didn't started to talk about Speed/Quality ;)

I think there are different approaches creating different look @ feels personally i like Infognitions Sharpness layer but i don't like how it sets the optimum for that keeping the aliasing out of the equation (in motion it looks as bad as spline36resize) :) and MotionDSP i like the aliasing layer but i don't like how it keeps the sharpness layer therefore out of the equation.

So my goal for Avisynth would be mixing both results in a balanced way together and im not sure if that even needs something computational heavy as needi2/3 at all but can be achieved much cheaper :)


Well, MotionDSP looks like a blurred spline36resize to me
Yep almost and that's i think a good way to start of from trying to improve the aliasing layer without using needi, and then reconstruct the sharpness layer (artificially or using temporal data) as best as possible :)

[Gser]

Topaz Enlarge2x which is supposed to be applied to a 2x nearest neighbor resize. Settings are default except Less de-noise is enabled.
http://forum.doom9.org/attachment.php?attachmentid=10981&stc=1&d=1272022158

[Motenai Yoda]

eedi3_rpow2(2,cshift="spline64resize")
mt_lutxy(binomialblur(1.0),yexpr="x x y - abs 1 x y - 10 / / atan * +",u=2,v=2)
http://img192.imageshack.us/img192/2862/asdadasdz.png

[2Bdecided]

The images from my digital still camera's video mode have far more aliasing than the original at the top of this thread.

Hence real super resolution would, in theory, give more benefit.

I haven't found any commercial algorithms that are anything like real super resolution - most aren't even as good as the anti-alias + sharpen we can do in AVIsynth.


btw, real super resolution works on multiple frames, so showing a screen cap from the very start of the sequence as an example (i.e. no previous frames) would disadvantage it, if we had it!

Cheers,
David.

[CruNcher]

exactly that's why i provided moving compressed examples of it and visually MotionDSP is a real super resolution you can clearly see that but it's a very hard one (but it was optimized for very destroyed compressed mobile sources) also Infognition is but in the contrary to MotionDSPs algorithm and settings used in Vreveal it's a soft one :)
Yodas result though looks nice from a visual standpoint, guess it could be also fine in Motion :)

[odditory]

pretty interesting stuff here. now i just need to get to grips with avisynth and be able to perform this on my own!

[Gser]

So Super-Resolution is more for frame restoration than resizing.

[CruNcher]

It can be used for several things it's versatile usable Deinterlacing, Denoising, Resizing and Detail Recovery comes to my mind

[NerdWithNoLife]

So next time I see a government agent on TV magnifying something 100x with lots of detail, I'll know it must have been nnedi! :devil:

[Stephen R. Savage]

This thread should be renamed "commercial consumer resizers vs avisynth solutions". None of these Avisynth examples involved any "super-resolution", which implies temporal averaging and motion compensation.

[2Bdecided]

This thread should be renamed "commercial consumer resizers vs avisynth solutions". None of these Avisynth examples involed any "super-resolution", which implies temporal averaging and motion compensation.None of the examples in this thread do, but there are at least two threads on doom9 where such scripts are posted.

Cheers,
David.

[CruNcher]

Yep i missed all of the discussion around it also your heated conversation with Didée about it :)
though you talked about flicker you see in your clips without using real super resolution, do you mean with flicker left aliasing in motion ?
All of the results i made show this except MotionDSP but therefore it's a blurry mess in the endresult which would indicate that there isn't enough aliasing in the source to begin with to Superresolute, though for me it rather looks like it's absolutely untuned for this kind of sources and expects hell of destroyed ones (though imho it still does a nice job on the anti-aliasing).
And yeah for Resizing it seems not really better except you have very aliased source and most of the times you gonna find this with Interlaced material or where the sensor where of low quality exactly the case for what MotonDSP is optimized for, also you can see it does a lot temporal repairing by default comparable visually would be i guess if you use repair() in Avisynth for the same effect.

here are some interesting links to read

http://forum.doom9.org/showpost.php?p=1386328&postcount=2
http://forum.doom9.org/showpost.php?p=1372215&postcount=4
http://forum.doom9.org/showpost.php?p=1372234&postcount=7
http://forum.doom9.org/showpost.php?p=1372791&postcount=12

Didées "Yet another Sharpening" looks interesting :) i wonder how it would do on the MotionDSP output ;P

[Motenai Yoda]

function HPblur(clip o) {
rg11=o.removegrain(11)
rg11D=mt_makediff(o,rg11)
xD=mt_lutxy(rg11D,rg11D.removegrain(11,-1),"x 128 - y 128 - * 0 < x y + 128 - x 128 - abs y 128 - abs < x y ? ?")
o.mt_makediff(xD,U=2,V=2) }

src=last
base=src.HPblur().HPblur()#.HPblur()
head=base.mt_adddiff(mt_makediff(base,base.removegrain(4,-1)))
shrp=base.mt_adddiff(mt_makediff(base,base.removegrain(11,-1)))
lim=shrp.repair(head,1)
shrpL=src.mt_adddiff(mt_makediff(lim,base),U=2,V=2)
shrpLL=shrpL.mt_adddiff(mt_makediff(lim,base),U=2,V=2)
shrpLLD=mt_makediff(src,shrpLL)
shrpLLx=src.mt_lutxy(shrpLLD,"x y 128 - abs 15 / 1 2 / ^ 15 * y 128 - 2 ^ y 128 - 2 ^ 2.3 + / * y 128 - y 128 - abs 0.0001 + / * -",U=2,V=2)
shrpLLxD=mt_makediff(shrpLLx,src)
backglow=shrpLLx.mt_makediff(shrpLLxD.removegrain(11,-1).removegrain(11,-1),U=2,V=2).mt_makediff(shrpLLxD.removegrain(11,-1),U=2,V=2) .mergeluma(shrpLLx,0.618)
shrpLLx # or backglow

[Keiyakusha]

To me it looks like last.Sharpen(0.7), except that it has less uniform effect

EDIT: and on a bit blocky sources this function leaves artifacts on the edges of the blocks.

[Undead Sega]

No examples from Video Enhancer? :D

[AnnaFan777]

good, more comparisons please.

[ReinerSchweinlin]

Here is a screenshot of vreveal. But since it´s the first frame and vreveal doesn´t supprt free image sizes, it´s not much use for comparrison. A videosequence with some steady background would show better the benefits of real super resolution.

http://www.abload.de/img/screenshot71windowstaseeuw.png

[AnnaFan777]

this vreveal / motionDSP is so blurry , I can't believe it. how dare they sell this garbage?

[ReinerSchweinlin]

vreveal just popped up with V3 - Did anybody take a look at it already?

[cweb]

vreveal just popped up with V3 - Did anybody take a look at it already?

I bought it and will be trying it.
My experiences with v2 were good, not blurry at all. In some cases I got better deshaking with it than with avisynth and I found it useful for that reason. Like always, YMMV.

[ReinerSchweinlin]

I bought my update, too, just a few seconds ago. The new panorama-Video Funktion seems nice and using my ATI will be fun - hopefully.

[ReinerSchweinlin]

Oh well, OpenCL needs Win7, should have checked in advance :)
The new GUI of Vreveal is nice, DIVX Support was cut out (only H264 with almost no options, MJPEG and unkomplressed AVI are left) and just tossing the sample clip at it:
http://www.abload.de/img/vrevelatest00002chb.png
Either I am missing something or this simply does nothing (at least when tlaking about super resolution capabilities, the other filters are usable, but that´s not the topic here).

[cweb]

Oh well, OpenCL needs Win7, should have checked in advance :)
The new GUI of Vreveal is nice, DIVX Support was cut out (only H264 with almost no options, MJPEG and unkomplressed AVI are left) and just tossing the sample clip at it:
http://www.abload.de/img/vrevelatest00002chb.png
Either I am missing something or this simply does nothing (at least when tlaking about super resolution capabilities, the other filters are usable, but that´s not the topic here).
Oh that's bad. I checked and yes I can confirm they reduced the output file formats.

I wrote to them, perhaps you might do so too, hopefully they will add huffy, divx etc.

[ReinerSchweinlin]

I had some nice talk. Unfortunately, they cut out super resolution suppert, so vreveal now "only" is a video-tool for deshaking, denoising, etc....

I´ve been reading here:

http://www.psi.toronto.edu/~vincent/videoepitome.html

The results look.... unbeleivable... Does anybody have any insight or tried out some of these scripts?

[Didée]

It doesn't seem surprising that a NN works well on the scene on which it has been trained.

[ReinerSchweinlin]

true. but still....

[Undead Sega]

I had some nice talk. Unfortunately, they cut out super resolution suppert, so vreveal now "only" is a video-tool for deshaking, denoising, etc....

I´ve been reading here:

http://www.psi.toronto.edu/~vincent/videoepitome.html

The results look.... unbeleivable... Does anybody have any insight or tried out some of these scripts?

I remember this! I even emailed him 6 years ago! and even up to now this method isnt been used.

[djonline]

http://www.infognition.com/articles/video_resize_shootout.html
I found some PSNR bench are wrong on this page. E.g., Avatar film.
Instant HD and Smart Edge are better for me than Video Enhancer, but .. they have position offset by 1px left and above. So PSNR is wrong for them.

[Didée]

Once upon a time, there was a shootout of free-energy-machines.

Years later it was found that the measurements for competitians D and E were wrong. They didn't consume 3000% more energy than they outputted, but only 2000% more.

An exceptional error that was. But doesn't change anything about the fact that there are no free-energy machines.

[2Bdecided]

You've got to lose this cynicism Didée*. Just because it's not magic does not mean it's not an improvement - and one where the various possibilities deserve to be compared as fairly as possible.

* - actually, don't - it's entertaining, and we need someone to call BS for what it is.

[wonkey_monkey]

http://www.infognition.com/articles/video_resize_shootout.html

Bicubic is "better" than NNEDI3?

David

[Forensic]

I have extensive experience with Ikena (the forensic version of vReveal) and prefer the results of avisynth scripted solutions. However, it would be an asset if MotionDSP released their SR functionality as a plug-in so it could be incorporated into avisynth/vapoursynth. It is my understanding that MotionDSP has shifted their focus toward real-time military applications.

[kolak]

No examples from Video Enhancer? :D

http://s7.postimage.org/i2oy8xbm3/image.png

Frame 12- 320x180 scaled 2x.

http://s4.postimage.org/x1wpoi7u5/srnn.png

Best of both worlds- Video Enhancer+NNEDI3- sharp and no aliasing.

[matfra]

Hey Kolak,
I like your last screenshot.
Can you provide me the AVISynth script so I can try it ?
Also, you said that its Video Enhancer+NNEDI3- sharp and no aliasing.
I thought that Video Enhancer NNEDI3 are doing the same thing. Upscaling. Why would you upscale 2 time a video ?

[kolak]

Video Enhancer upscale than:
NNEDI3(field=-2)
Merge(SelectEven(),SelectOdd())

removes vertical aliasing with not that big loss in sharpness (you can sharpen at the end).
Just note that Video Enhancer does not shine that much on real sample (eg PAL to HD upscale) as on this original_res/2 downscale than 2x upscale test.

[kolak]

30 days trial:

http://www.infognition.com/VideoEnhancer/

there is no free version. If you like it you need to buy it.

Use google for search- it's quite easy :)

[matfra]

The program cost about 30$ now. What do you think about it ?
I just tried it. Its nice, and run fast at Speed Settings. Im not sure the Quality Setting give a better result than a custom Avisynth script ?
How Video Enhancer perform against other Avisynth Upscaler ? Does Video Enhancer perform a lot better than NDDI3 ?

[kolak]

Try yourself- the best what you can do to get answer for your question.

[laserfan]

Not $30, but $40 afaict. I wonder kolak if you (or anyone) can explain the difference between the VideoEnhancer application and the Super Resolution VDF plugin (they charge $40 for either one). Is it just that the app is simpler to use, vs. Vdub with the SR plugin i.e. would be more for experts?

[Rumbah]

Well, before you buy it you should use the trial version to test it for your self with a real sample (not a downscaled version).

In my testings 2 years ago it didn't look better than a NNEDI upscale with light sharpening. But I admit it's an easy to use all in one thing.

[kolak]

Not $30, but $40 afaict. I wonder kolak if you (or anyone) can explain the difference between the VideoEnhancer application and the Super Resolution VDF plugin (they charge $40 for either one). Is it just that the app is simpler to use, vs. Vdub with the SR plugin i.e. would be more for experts?

As far as I understand there is no difference- one is just standalone app another plugin for Vdub. I think output is the same and you can use the one which suits you better.

[matfra]

Is there a way to run Video enhancer in Avisynth script ?

[kolak]

No idea- don't think so.

Please start a new thread if you want to discuss Video Enhancer itself. This is going out of topic.

[laserfan]

Well, before you buy it you should use the trial version to test it for your self with a real sample (not a downscaled version).

In my testings 2 years ago it didn't look better than a NNEDI upscale with light sharpening. But I admit it's an easy to use all in one thing.

As far as I understand there is no difference- one is just standalone app another plugin for Vdub. I think output is the same and you can use the one which suits you better.
Thanks for the feedback. My main interest has always been improving DVDs but these things seem aimed instead at even lower resolutions than DVD...

In any case I'll look at NNEDI before popping for $40 bucks. :scared:

[matfra]

Please start a new thread if you want to discuss Video Enhancer itself. This is going out of topic.

Lol, No need to be rude Dude !