Log2Lin's lin2log math might be incorrect

[markreidvfx]

Hi,

I was looking at the implementation of log2lin for reference and I think the lin2log math might be incorrect.

https://github.com/NatronGitHub/openfx-misc/blob/master/Log2Lin/Log2Lin.cpp#L146

the comment there says

offset = pow(10,(blackpoint - whitepoint) * 0.002 / gamma)
gain = 1/(1-offset)
linear = gain * (pow(10,(1023*v - whitepoint)*0.002/gamma) - offset)
cineon = (log10((v + offset) /gain)/ (0.002 / gamma) + whitepoint)/1023

To me, the should cineon lin2log expression be

cineon = (log10(v/gain + offset )/ (0.002 / gamma) + whitepoint)/1023

v/gain + offset instead of (v + offset)/gain

Also the comment shows up twice in the file and the first one is different

[devernay]

There is definitely an inconsistency between the two formulas.
Note that TuttleOFX has the same issue:

• https://github.com/tuttleofx/TuttleOFX/blob/36fc4cae15092a84ea8c29b9c6658c7cabfadb6e/libraries/terry/src/terry/colorspace/gradation.hpp#L348
• https://github.com/tuttleofx/TuttleOFX/blob/36fc4cae15092a84ea8c29b9c6658c7cabfadb6e/libraries/terry/src/terry/colorspace/gradation.hpp#L381

I can't find where I got these formulas from. Do you have a reference? Maybe the log2lin formula is wrong, and not the lin2log?

[devernay]

This doc, which is to be the foundation to Cineon, seems to imply that log2lin is wrong, rather than lin2log:
http://www.dotcsw.com/doc/cineon1.pdf

[devernay]

Does this look good to you? c89ac63

[markreidvfx]

The inverse math and the spec make sense to me!

I thought the lin2log was wrong because the log2lin matches nuke.

But the spec implies nuke's math is wrong

(pow(10,(1023*x-685)/300)-.0108)/(1-.0108)

they shuffled a couple things around, divides instead of multiplies

1 /0.002 / 0.6 == 300
offset == 0.0108

The minus offset is happening before the gain.

am I crazy?

[devernay]

@markreidvfx I see that python colour' version refers to This formula in Sony's code which implies that lin2log is wrong, as you initially said. Do we agree on this? (sorry, I need you as a reviewer)

Edit: There's also this code in ACES

[markreidvfx]

Yes I agree, I think it might be best fix the lin2log like I initially suggested. python colour's lin2log version seems to also match the one I formulated by inverting the log2lin.

I still find it a bit un-nerving that the spec says otherwise, but I guess it is better to be compatible other implementations.

[devernay]

Does #79 look right?

[devernay]

closed by #79

[markreidvfx]

sorry, I missed this, looks good to me!