Upscaling Newbie Question

[bxyhxyh]

It is said upscaling DVD don't magically create new details.

What about nnedi3_rpow2().Debilinear() combination?

[Asmodian]

Debilinear() is not meant for that, debilinear is for undoing a previous bilinear upscale. Google tells me some people use it on the chroma; they are assuming the studio used a bilinear resize to get the 4:2:0 chroma from a 4:2:2 or 4:4:4 master but that seems crazy to me, especially with a DVD source.

NNEDI3 can make up details, sort of, but they aren't the real details. If I wanted to make upscaled DVD rips I would use NNEDI3_rpow2() followed by a downscale to the desired resolution using my favorite linear light resizer (I like Dither_resize16nr from cretindesalpes' dither tools (http://forum.doom9.org/showthread.php?p=1386559)). The inverted kernels (debilinear and similar) are not appropriate for this step as the output of NNEDI is not from that kernel.

[feisty2]

if u really wanna blow up sd clips to hd, a focus deconvolution filter + edi resizer should do the job

[bxyhxyh]

Yes, debilinear is not for this. But when I use it on good source, result looked good.
So I thought it can be good choice to upscale this way.

http://check2pic.ru/compare/36221

[Asmodian]

What about nnedi3_rpow2().ResampleHQ() or:

nnedi3_rpow2()
Dither_y_gamma_to_linear()
Dither_resize16nr()
Dither_y_linear_to_gamma()


What did you use for "upscale"?

[feisty2]

No, downscaling with dither_resize16nr makes no sense cuz downscale ringing is almost invisible, so it will only bring u blur and other troubles and give u no good

[bxyhxyh]

For this case downscaling with Debilinear does strengthen the details, also strengthen the artifacts, like Didée's FineSharp.
It didn't create ring by itself, except blacks bars were not removed before Debilinear.

Isn't it more effective than nnedi3_rpow2().someresizer().somesharpener() ?
At least it has less rings.

[Asmodian]

No, downscaling with dither_resize16nr makes no sense cuz downscale ringing is almost invisible, so it will only bring u blur and other troubles and give u no good

Good point. Should drop the nr for downscaling and the to_linear and back for upscaling. :o

For this case downscaling with Debilinear does strengthen the details, also strengthen the artifacts, like Didée's FineSharp.
It didn't create ring by itself, except blacks bars were not removed before Debilinear.

Isn't it more effective than nnedi3_rpow2().someresizer().somesharpener() ?
At least it has less rings.

I must admit upscaling DVDs isn't something I do. Do you really need more sharpening than NNEDI3 gives you?

Debilinear in not necessarily a better way to sharpen than someresizer().somesharpener() but if you like it go for it. ;)

[feisty2]

Debilinear is something like a resizer + a deconvolution filter, all deconvolution filters will increase sharpness and that's why u feel the details are enhanced, but debilinear isnt the right deconvolution filter for upscaling, u wanna increase overall sharpness for upscaling? a focus deconvolution kinda filter should do a far better job

[bxyhxyh]

Thanks guys for your advises and explanations.
a focus deconvolution kinda filter should do a far better job
Good deconvolution filters are?

[raffriff42]

Here is one I like: NonlinUSM (http://forum.doom9.org/showthread.php?p=1555234#post1555234)### enhance contrast - like unsharp
## by Didée
## http://forum.doom9.org/showthread.php?p=1555234#post1555234
##
## @ z - default 6 (3 = subtle; 16 = massive)
## @ pow - default 1.6 (4.0 = mild; 1.0 = oversharp)
## @ str - default 0.7 (3 = strong; 18 = massive)
## @ rad - default 9 (3 = fine detail; 18 = soft)
## @ ldmp - default 0.001 (not tested)

## enhance detail in upscaled, high quality source
NonlinUSM(rad=1.5, str=1.0).Sharpen(0.3)

## restore some low-level detail in low bitrate source
# http://forum.doom9.org/showthread.php?p=1667813#post1667813
NonlinUSM(pow=4)

[feisty2]

I donno if any focus blur deconvolution filter for avisynth already existed or not, but I plan to develop one
here's one before/after comparison
before
http://thumbnails109.imagebam.com/31534/924cbf315331433.jpg (http://www.imagebam.com/image/924cbf315331433)
after deconvolution (blur radius=1.2)
http://thumbnails109.imagebam.com/31534/ba9a66315331450.jpg (http://www.imagebam.com/image/ba9a66315331450)

it's still buggy, I'm testing it and maybe someday I'll release it

[feisty2]

for now, u can use "sbr" as ur sharpen core to produce similar result like deconvolution above

[raffriff42]

wat is "sbr"

(I tried FQSharp (http://www.avisynth.nl/users/vcmohan/FQSharp/FQSharp.html)but did not like it; too ringy. Probably doing it wrong.)

[feisty2]

here http://nmm.me/yy
"GSMC_sbr" a supporting function for "GrainStabilizeMC" by mawen1250
I donno where's it originally from, don't ask me :P
I ain't used FQSharp before, I donno why it's ringy, do u have the source code of FQSharp?
however my deconvolution filter doesn't seem so ringy unless u use a very large radius, but I'm still developing it, it's just way too buggy to be released now
maybe I'll release it someday