Normalize (-1,1) in Designer ?

Eric Chadwick

Looking for a way to process a 24bit texture such that the red and green channels will be normalized within a -1 to 1 range, and not include the blue channel. 

What does the Normal Normalize node do? I guess it’s only normalizing 0-1 ?

This is intended for an anisotropy texture. But I guess the same process would be needed for flow maps too.

The blue channel is not used for the vectors, it’s a mask for strength of the effect, so I wouldn’t want that factored into the normalization.

Eric Chadwick

What does the Normal Normalize node do? I guess it’s only normalizing 0-1 ?

Answering my own question here. Ctrl E on the node:


So, first it normalizes to -1,1 then it converts that into 0,1. I guess I just need to make my own custom pixel processor, and simply omit that last part?

poopipe

Yep, exactly

Make sure your pixel processor is set to absolute color depth of 16f or 32f or it will clamp all the values In the function between 0 and 1.

Eric Chadwick

Hmm, I’ll need to check that when I get back to my PC. 

I posted my results here for feedback from the rest of the glTF developers https://github.com/KhronosGroup/glTF/issues/2290

Eric Chadwick

I’m trying to visualize what a -1,1 normalized image should look like in RGB. 

Ignoring the blue channel, since this anisotropy texture is XY vectors only. We can’t represent negative red values in 8bit RGB. So what would the result look like? 

poopipe

its just re-ranged to 0:1 and will be expanded back out when it's sampled by the shader - exactly the same as a normal map
just divide the resultant [-1:1] vector by 2 and add 0.5 

Eric Chadwick

So I guess I can't store the png bitmap normalized to -1,1 since the negative range gets clipped. Unless I use a 16 bits per channel PNG? The engineers are asking for a -1,1 normalized texture, instead of a 0,1 texture which I'm more familiar with. I guess they're really just asking for the latter, and will convert it to the former in the shader?

poopipe

Looking through the spec suggests png can theoretically support signed values but I don't think I've ever seen it done and I'd be surprised if most exporters support it. you'd probably want to use tiff or exr (possibly .HDR)  as a source format to store the numbers. it'd need compressing to a gpu friendly format that supports negative values as well
bc6h can support signed values - its not small though (i think 4x the size of dxt1 but possibly 8x)

TLDR : just don't

this should really be treated the same way as a normal map,  store the values in 0:1 range, expand it in the shader.

Eric Chadwick

OK that's what I was hoping. Thanks for the gut check!

Eric Chadwick

New question... is this a good solution? 

I want to normalize a 2D vector. The vector is stored in the Red and Green channels of a 24bit PNG. I want to pass through Blue and Alpha unchanged. I want to publish the SBSAR on GitHub for others to use in whatever app they use which supports SBSARs.

Any improvements I could make here?

Eric Chadwick

This seems to work well.

Although when I attempt to Bake Outputs via a Substance2 node in 3ds Max, it ignores the alpha channel.
Going to test with Substance Player.

Eric Chadwick

Thinking about adding a Boolean switch for 8bit vs 16bit output.

Is there a way for the SBSAR to auto-detect the resolution of the incoming bitmap and set that as the default resolution for the output? Same for bit depth?

poopipe

I'm pretty sure you have to specify an output size if the graph is relative-to-parent - I'd expect the same to be true for bit depth.
there's not really anything useful you can do to simplify the pixel processor 
you only need the pixel processor though, the rgba split/merge stuff is redundant cos the pixel processor is passing B(z) and A(w) out unchanged

Eric Chadwick

Thanks for this. I’ll remove those nodes. 

Are there any options for dithering to 8bit vs not?

Thinking I should probably set the graph to 32f from the start, not restrict to 8bit. So then you could feed a 32bit or 16bit or 8bit source.

Anyhow thanks again.

poopipe

the specific output settings should be handled by the plugin although if you set absolute parameters (eg. bit-depth/dimensions) in the graph they will carry through assuming the plugin supports them. 

Only the pixel processor itself needs to be 16f/32f in this case - what goes in and out can be handled as normal 8/16bit data because it's in the 0:1 range. 
You could (in theory) safely leave the graph as relative to parent - I've had bugs in the past with this stuff failing and had to force 16bit on .sbsar but it's been a few years since it last came up. 
fwiw - i haven't used 8 bit textures for years, the DXT compressed version of a 16bit image is no bigger than an 8bit one and you dont have to fight banding

Eric Chadwick

Thanks. 

No dithering settings, hmm.

Can’t use DXT since this is glTF, so JPG or PNG or KTX2. I’ll try the latter with 16bit.

Eric Chadwick

Finished it up! Thanks for all the help.
https://github.com/KhronosGroup/glTF-Sample-Assets/blob/1d4b083ca989bbf594309cb2ea66b4dc89a84783/Models/AnisotropyBarnLamp/README.md#normalizing-anisotropy