Adaptive sharpen - Image sharpening shader

[lanzorg]

Originally Posted by bacondither

New version 2015-09-23 released.

Tried optimizing some, it went faster for AMDs old VLIW gpus but slower for newer GCN. So that did not make it to this version.

The bilateral blur from the experimental version did better sometimes and worse other times. But the difference was so minimal and it was slower. So it's not included now.

Could you port this version to glsl?

[ts1]

glsl version here
This shader also requires floating point processing apparently.

[vood007]

Originally Posted by ts1

Here is glsl version of this shader:
Pass 1
...
This shader also requires floating point processing apparently.

Thanks for this. May i request a glsl version of the old one pass version too? Its linked in the first posting.

[ts1]

glsl version of Adaptive Sharpen

[vood007]

Originally Posted by ts1

vood007

// Copyright (c) 2015, bacondither
// All rights reserved.
...

Thanks again. Works fine under Linux with MPV.

[bacondither]

Originally Posted by madshi

@bacondither, can you double check AdaptiveSharpen with this test image?

http://madshi.net/FineSharpTest3.png

Last time I checked, AdaptiveSharpen makes this image brighter. Which is what most sharpeners do, so it's not surprising. But ideally, that should not happen. Maybe you can work on that? E.g. by using linear light in some parts of your algorithm? Or maybe by running a second pass which "repairs" modified brightness?

Thanks!

I agree that in some situations local contrast can be increased too much, but sometimes it should get brighter.
For example, if you blur the example image with a gaussian filter or a median filter it gets darker, "unblurring" it then should then make it look brighter.
But the original example image is pretty sharp already, increasing the local contrast/acutance a lot does not look that good in that example.

I have not found a satisfactory solution yet that does not compromise other situations. Sometime in the future i'm going to try testing a third "repair" pass.

I'm pretty sure i'll get some other ideas while in the shower or something...

Originally Posted by ts1

glsl version here
This shader also requires floating point processing apparently.

Just had a quick look, i can't understand why is the inputs into tanh(x) is clamped to -+20. That could lead to some screwy results.

[burfadel]

oooh new test version, thanks

[ts1]

That is not my port but one of mpv's developers. We made ports in the same day

[bacondither]

Originally Posted by ts1

That is not my port but one of mpv's developers. We made ports in the same day

Ah, okay.

I don't have MPV ready for testing now, but did some tests on it a while ago on ubuntu. Could not get some script for switching display frequencies to work as i wanted.

[ts1]

Windows builds available also. Just create "portable_config" folder next to mpv.exe with your shaders and mpv.conf with content something like:

vo=opengl-hq:fbo-format=rgba16f:post-shaders="~~/adaptive-sharpen-pass1.glsl,~~/adaptive-sharpen-pass2.glsl"

[madshi]

Originally Posted by bacondither

For example, if you blur the example image with a gaussian filter or a median filter it gets darker

Does it, though, when doing the blurring in linear light? Which is how blurring should ideally be done.

[bacondither]

Originally Posted by madshi

Does it, though, when doing the blurring in linear light? Which is how blurring should ideally be done.

Yes, forgot that my image editor did blurring in gamma light. Blurred it in linear light and the apparent brightness was then about the same as the original.

Did some tests and saw that doing the summation of the negative laplace vales in "fast" linear light did reduce the brightening effect.
Have not seen any negative effects of doing it that way yet.

Added it to the current experimental version.
Experimental - 2015-10-29

[madshi]

Great, thanks - any progress is welcome!

One other thing that I've noticed when comparing FineSharp with AdaptiveSharpen: They have a quite different look. FineSharp somehow makes lines (at least dark lines) appear thinner, while AdaptiveSharpen seems to have a certain tendency to make dark lines fatter. Personally, I like the look of thinner lines. Is that something that could maybe be improved in AdaptiveSharpen?

[aufkrawall]

Originally Posted by bacondither

Added it to the current experimental version.
Experimental - 2015-10-29

Amazing, image brightness pretty much doesn't increase and it still looks as good as usual otherwise.

Yes, line thickening is noticeable, especially with cartoons. This can look nice with low strength after soft scaling algorithms, but with higher strength it gets a bit too obvious.

[jewshawn2]

I can't seem to get any of the new 2 pass shaders to work with potplayer. However, the early single pass shader works fine.

I use the "Shader Stacker" but no matter what I do I end up with a green screen. But MPC-HC works perfectly fine.

I would much appreciate it if someone else can someone verify?

[toniash]

Originally Posted by jewshawn2

I can't seem to get any of the new 2 pass shaders to work with potplayer. However, the early single pass shader works fine.

I use the "Shader Stacker" but no matter what I do I end up with a green screen. But MPC-HC works perfectly fine.

I would much appreciate it if someone else can someone verify?

In "trade quality for performance" you must UNTICK:
store custom pixel shader in 16 bit ...
use 10bit chroma ...
use 10bit image ...
run custom pixel shaders in video levels ...

hope it helps

[CruNcher]

Not sure which version is used inside of MPC-BE currently but it does a great job at relatively low shader costs (~10% above the simpler ones)

Very impressive results even on rather complex textures but shimmering in high frequency skylines for example still seems to be a big problem at least in the version used inside of MPC-BE.

I use it as a very decent post resize pass solution, which works out for most different compressed MPEG content just great without oversharpen it to much that it feels odd

[bacondither]

Originally Posted by jewshawn2

I can't seem to get any of the new 2 pass shaders to work with potplayer. However, the early single pass shader works fine.

I use the "Shader Stacker" but no matter what I do I end up with a green screen. But MPC-HC works perfectly fine.

If you use madvr as renderer you could also try unticking the "lose BTB and WTW if it improves performance"

Originally Posted by CruNcher

Not sure which version is used inside of MPC-BE currently

I think it uses the old one-pass 2015-06-25 one.

I could backport some things to the old one pass version and make it better for a very small performance cost.

[bacondither]

The new 2015-11-05 experimental should be a little bit faster then the 2015-10-29 one. ~1% to 2%.

[vood007]

Originally Posted by bacondither

I could backport some things to the old one pass version and make it better for a very small performance cost.

This would be awesome! I still use the older 1-pass version because it allows me to stay at the lowest power-state of my GFX.