Idea & Info: Detail Preserving Upscale GOUT

I haven't been keeping up, is TeamBlu's upscale now available?

Still says "Coming Soon" on the project thread's first post.

Thanks, I had trouble finding the thread.

The best results I've seen for upscaling involve Avisynth's nnedi2_rpow2 and ResampleHQ (a gamma-aware downscaler). I'd certainly say it's "detail preserving".

Assuming a PAR of 72:79 (since we're using 711 rather than NTSC's 710.85), the GOUT has a DAR of ~64:27 (2.37:1) since 711*72/79=648 and GOUT has 274 visible scanlines, so 648/274=2.364963503649635...

The idea is to use nnedi2_rpow2's field interpolation to get from 720x480 to 1440x960. (This'll take care of many of the aliasing problems as well!)

Next, you use ResampleHQ to downsize to 1280x540 (or what-have-you) by using the extended parameters to take care of the cropping. It's better to pass the crop values to ResampleHQ rather than crop manually since ResampleHQ can make use of the pixels-that-will-be-cropped in its downscaling algorithm, thereby reducing the likelihood of screw-ups at the borders.

(Note: You'll have to double any crop values that you computed since you'll be running ResampleHQ on the double-size output from nnedi2_rpow2.)

If someone posts a frame (preferably a PNG) from GOUT here, I can hack the code up rather quickly.

When you look at this video, when he switches between the normal scaling and the "detail preserving upscale" you can see, that all this new function does is it pretty much just blurs the image to get rid of pixelation and then it applies an unshrap-mask kind of edge enhancement to the blurred image.

It is also clearly visible in the sample from emanswfan - the edge enhancement is pretty horrible.

As to Spaced Ranger's link, for me the clear winner is the "SAR DataDependentLanczos" which seems to strike the best ballance between sharpness and unnatural edge smoothing (and apparently it's freeware) - some of the filters make the image look really sharp without applying edge enhancement but they turn the whole image into large sharp colored blotches, so it starts looking like a painting. I'd say the second best is VideoEnhancer, which preserves the most detail but has some pixelation and some mild edge enhancement.

Edit: Actually, looking at it again, the VideoEnhancer may actually be preferable in many cases, since it really does preserve a lot of detail - the SAR DataDependentLanczos would totally rock for upscaling animation though.

It would be interesting to see what AntcuFaalb's method would do with that image.

And another edit: Before, I only looked at the image from Panasonic4.avi and now I looked at some of the others and the SAR DataDependentLanczos actually seems to create some really weird artefacts, which would probably look really bad in motion, so the VideoEnhancer totally wins.

I tried using nnedi3_rpow2 on the GOUT, but didn't get as good a result as DJ's team.

I don't think they ever posted the upscale script they were using, but I suspect it includes QTGMC in progressive-source mode.

GOUT looks much better on CRT displays, so when cretindesalpes came out with his new accurate crt_display Avisynth function I jumped on the opportunity to see how well it handles GOUT.

Wow. Just wow. See for yourself…

1. http://i.imgur.com/nK9jWnB.jpg

2. http://i.imgur.com/4f3VXGe.jpg

In general, it looks nice, but IMO I'm not sure he got what he was aiming for with the scanlines.  I'd put more of a glow around each scanline, to make the effect less obvious.  As it stands now, it looks a little too much like a digital display trying to be an analogue display (which, I guess, it is, but still...)

CatBus said:

I'm not sure he got what he was aiming for with the scanlines.

The scanlines weren't explicitly put in. His script actually emulates how an aperture grille CRT (and other CRTs) functions and the scanlines in this case are a side-effect of that.

JEDIT: Check out his other samples: https://www.doom9.org/showthread.php?s=3a7a37048d43d2750ea9abe6aa4dd0b0&p=1683051#post1683051

JEDIT 2: Also keep in mind that this is best viewed at a distance. Neither a CRT nor this look "right" when examined up close.

I've been thinking about this for a a little while now -

First off, with the CRT Emulation mode - is there any way to apply a pixel shift to every other frame, or in a particular pattern in avisynth?

I was thinking if you set a 4 frame pattern for this, you could then apply an interlace filter to the resulting video, as it would make the video appear interlaced (I could be way off, but thought it might work)

You would set the first frame in it's original position, the second shifted 1 frame up, the third frame back to the original position, the fourth frame shifted 1 frame down.  

I was also curious about using multiple upscaling methods, then finding the mean chroma and luma using the MergeChroma/MergeLuma functions. Doing this, you could attempt to find the perfect combo/middle ground between spline, bicubic, and lanczos, perhaps even throwing in a small amount of bilinear to smooth any over sharpening.

There is also a median method here: http://forum.doom9.org/archive/index.php/t-140282.html

Not sure about a mode function, but to me that almost seems the best possible way to mix the files, as it would only use the chroma/luma information that occured most frequently across the different upscales. This would hopefully eliminate some of the artifacting/haloing in lanczos.

I find this idea of trying to simulate a CRT display equal parts ridiculous and fascinating, so I decided to take a stab at it myself :-D

https://drive.google.com/file/d/0B0dtCY1O9WXzVmNKRU1TbWprWEk/view?usp=sharing

Harmy said:

I find this idea of trying to simulate a CRT display equal parts ridiculous and fascinating, so I decided to take a stab at it myself :-D

https://drive.google.com/file/d/0B0dtCY1O9WXzVmNKRU1TbWprWEk/view?usp=sharing

 It's almost as if I have my face too close to the television!

Here are a few screenshots:
http://s25.postimg.org/itfnsztal/Pre_comp_1_0_00_56_01.png

http://s25.postimg.org/ay0j0fua5/Pre_comp_1_0_01_38_00.png

http://s25.postimg.org/og7fcq6fh/Pre_comp_1_0_01_33_03.png

NeverarGreat said:

It's almost as if I have my face too close to the television!

You are in this case. Take a few steps back to see how well it upscales the image.

I guess it does increase the illusion of detail.

NeverarGreat said:

I guess it does increase the illusion of detail.

I just think it does a fantastic job of upscaling without losing detail.

AntcuFaalb said:

CatBus said:

I'm not sure he got what he was aiming for with the scanlines.

The scanlines weren't explicitly put in. His script actually emulates how an aperture grille CRT (and other CRTs) functions and the scanlines in this case are a side-effect of that.

That's not really accurate, at least it doesn't look that way to me. On a CRT you only really see scanlines in progressive mode (e.g. old video games, arcade machines, computer monitors). In interlaced mode the scanlines are offset by the following field, but the scanline itself is significantly thicker than it is for the empty space between the scanlines. Here, I'll make a picture...

Simple example (no RGB offset):

Top field:

Bottom field:

Combined:

And this is what happens when all you do is upscale (spline36resize) and then "add" scanlines (really you're adding the "black" between the scanlines):

The difference of course being that spline36resize would keep the film at its original speed-rate, whereas the combination of fields would produce a double-rate file (50 or 59.97 fps) since each field is visible for two frames. You could certainly design such a script that leaves you no visible darkened lines and retains the correct brightness of the image; which is more akin to how it looks on a CRT.

Anyway, yeah you could do that, leave yourself a more solid looking picture, and then apply the RGB offset, but it still won't really look the same as viewing a CRT - better to buy yourself a cheap CRT tellie and watch it direct!