Super xBR upscaler

[luquinhas0021]

What are the parameters of this algorithm? How I choose the number of taps? This thread was created for discussing this method.

[MysteryX]

Look within the super-xbr.avsi (https://github.com/mysteryx93/AviSynthShader/blob/master/Shaders/Super-xBR/super-xbr.avsi) file to see how the parameters work, or read the documentation (https://github.com/mysteryx93/AviSynthShader/blob/master/Release/ReadMe.txt).

Or you can also try it with Media Player .NET or madVR during live playback as it is integrated in those.

[Hyllian]

What are the parameters of this algorithm? How I choose the number of taps? This thread was created for discussing this method.
The number of taps can't be changed in current implementation. It uses 2-taps, though it isn't limited to it, so if you wanna improve it, you can tweak the code to use more taps.

Anyway, I wrote a small text (https://drive.google.com/file/d/0B_yrhrCRtu8GYkxreElSaktxS3M/view?usp=sharing) very friendly trying to explain the concepts behind Super-xBR a while ago.

[MysteryX]

There's also something wrong with the current implementation. Compared to MPDN, it is slightly shifted to the right and slightly stretched vertically. Not sure what's wrong.

http://s20.postimg.org/6dnyr49g9/xbr_MPDN.jpg (http://postimg.org/image/6dnyr49g9/) http://s20.postimg.org/amxqqej15/xbr_Avi_Synth.jpg (http://postimg.org/image/amxqqej15/)

Current implementation is here (https://github.com/mysteryx93/AviSynthShader/tree/master/Shaders/Super-xBR)

[Hyllian]

There's also something wrong with the current implementation. Compared to MPDN, it is slightly shifted to the right and slightly stretched vertically. Not sure what's wrong.

Current implementation is here (https://github.com/mysteryx93/AviSynthShader/tree/master/Shaders/Super-xBR)

Madshi has fixed it before and I forgot to update my own implementation. To solve it, you have to add a line at the third pass after these two lines:


float2 g1 = float2(1.0/IN.texture_size.x, 0.0);
float2 g2 = float2(0.0, 1.0/IN.texture_size.y);
VAR.texCoord -= 1.0/IN.texture_size; // Add this!


I don't know how you'll do this at the MPDN implementation as I'm not aware of it.

It can't be fixed for the 2-pass version, though. So, I think Super-xBR works much better by using 3-passes.

[MysteryX]

Why is MPDN's version working then?

I'm not finding those lines
https://github.com/zachsaw/MPDN_Extensions/blob/master/Extensions/RenderScripts/Super-xBR/super-xbr.hlsl

Edit: oh, it appears my version is outdated!
https://github.com/mysteryx93/AviSynthShader/blob/master/Shaders/Super-xBR/src/super-xbr.hlsl

[MysteryX]

Oh shit... this latest update seems to be PixelShader model 5!

but MPDN, as well as madVR, can run either on DX9 or DX11. How do they run these on DX9?

[Hyllian]

I have fixed it in my implementation: https://github.com/libretro/common-shaders/tree/master/xbr/shaders/super-xbr

You need to find the main texel coordinates and subtract a pixel_size of the texture being filtered. That's what that line do. It fix the pixel shift. I don't know how you can do this in these other implementations.

[MysteryX]

His implementation looks about the same as yours, but he split the passes to be compiled separately.

Either way, it's Shader Model 5 which requires DirectX 11

[luquinhas0021]

Why worry about pixel shift, and have the work of update the MPDN Super-xBR, if the shift difference between MPDN Super-xBR and AviSynth Super-xBR is so little, but... Avisynth Super-xBR is notorious sharper than MPDN super-xBR?

[sqrt(9801)]

Edit: oh, it appears my version is outdated!

It isn't. I was using the current version from MPDN when I made my AVS script.

I am not sure where the shift/stretch come from, though. I am fairly certain they were not there when I made my tests (with third pass enabled), although I've only tried the shader on RGB32 clips.

[MysteryX]

It isn't. I was using the current version from MPDN when I made my AVS script.
This is the current version in MPDN, which was updated on June 27th. Not sure where you got the PS3 version from.
https://github.com/zachsaw/MPDN_Extensions/blob/master/Extensions/RenderScripts/Super-xBR/super-xbr.hlsl

My version is a bit sharper simply because the settings were slightly different in both screenshots.

Updating the code to work with DirectX 11 wouldn't be difficult. I just need to figure out whether DX11 supports Shader Model 3. Then the hardest part would be to change the way parameters are being configured.

[sqrt(9801)]

It's the same file. I've just removed the structs and the vertex shader code.

[MysteryX]

It's the same file. I've just removed the structs and the vertex shader code.
Oh, so you ported it. Would you be able to port the fix; or port a version that is actually working well?