MrLixm/AgXc

[OCIO] ACES interchange inaccuracy

Closed this issue · 6 comments

conlen-b commented

The conversion to ACES2065-1 for interchange seems to be incorrect-- I found this error originally with the following process:
I have a scene-referred render from Maya + V-Ray with the Linear sRGB rendering space from your config. I bring this into a Nuke scene set up with your OCIO config with a read node with the IDT set to Linear sRGB. Using an OCIOColorSpace node, I convert Linear sRGB -> ACES2065-1. I write this out to a 16bit EXR using a write node with the ODT set to Passthrough.

In a different nuke scene using the cg-config-v1.0.0_aces-v1.3_ocio-v2.0.ocio OCIO config from https://github.com/AcademySoftwareFoundation/OpenColorIO-Config-ACES/releases, I read in the 16bit EXR from the previous Nuke scene with the IDT set to Raw. With a second read node, I read in the original scene-referred render with the IDT set to Raw. I then use an OCIOColorSpace node to convert the second read node from Linear Rec.709 -> ACES2065-1. If I set my viewer to Raw (sRGB - Display) and compare the two, they are close but still pretty off. Liam was able to recreate this issue as well.

MrLixm commented

Thanks for reporting, manage to reproduce the issue on my side.
Will investigate what's going on

MrLixm commented

Just created a new branch with a new module that allow me to reproduce the generation of the MatrixTransform I used in the OCIO config :

https://github.com/MrLixm/AgXc/blob/16-ocio-matrix-diff/dev/ocio/colorspaces.py

Executing this module gives me :

xyz_to_ap0()=[1.0604675987, 0.008384430217, -0.014338952537, 0.0, -0.486910537062, 1.370790753303, 0.093070834934, 0.0, -0.005306229727, 0.006434186849, 0.916879121543, 0.0, 0.0, 0.0, 0.0, 1.0]
srgb_to_xyz()=[0.4124, 0.3576, 0.1805, 0.0, 0.2126, 0.7152, 0.0722, 0.0, 0.0193, 0.1192, 0.9505, 0.0, 0.0, 0.0, 0.0, 1.0]
ap0_to_srgb()=[2.526838601825, -1.115756102267, -0.368990640029, 0.0, -0.297761286039, 1.376149938024, -0.084965677227, 0.0, -0.009234528688, -0.171078993092, 1.168621917667, 0.0, 0.0, 0.0, 0.0, 1.0]

The 2 first matrix corresponds to what I have in the config.

The last matrix is suppose to correspond to what is used in the ACES cg config :
https://github.com/AcademySoftwareFoundation/OpenColorIO-Config-ACES
For the Linear Rec.709 (sRGB) colorspace.

Though in the config here is the matrix specified :

[2.52168618674388, -1.13413098823972, -0.387555198504164, 0, -0.276479914229922, 1.37271908766826, -0.096239173438334, 0, -0.0153780649660342, -0.152975335867399, 1.16835340083343, 0, 0, 0, 0, 1]

Which as you can notice is not the same as what I mange to produce with my method.

Next step find why.
I think I need to check if my code is using the official matrix (with rounding issues) or the derived one for the primaries.

MrLixm commented

Found the issue, it was of course in my matrix generation code.
I mixed up the order in which I must concatenate the CAT matrix + I should also have used derived matrices (at least that what are using the official ACES configs)

You can check the fix in commit 61f076a

I will now propagate it to the OCIO config

MrLixm commented

Actually no, I did not fix anything, the issue is still here as that new commit still produce the same matrix.
I manage to find that the issue is on the CIE XYZ D65 to AP0 matrix (also to AP1).
The sRGB > CIE XYZ D65 matrix is a perfect match with the ACES config.

  • 👍1
MrLixm commented

I finally have it, the order is which is was multiplying the matrices when going from RGB to XYZ was not correct, I had to inverse it.
Fix in commit 8c9a140

For confirming it was working I used :

import colour
import numpy

d65 = colour.CCS_ILLUMINANTS["CIE 1931 2 Degree Standard Observer"]["D65"]
ap0: colour.RGB_Colourspace = colour.RGB_COLOURSPACES["ACES2065-1"]
ap0.use_derived_transformation_matrices(True)
array = numpy.array([0.5, 0.4, 0.3])
converted = colour.XYZ_to_RGB(
    array,
    d65,
    ap0.whitepoint,
    ap0.matrix_XYZ_to_RGB,
    "Bradford",
)
print(converted)

matrix = numpy.array( [ [...], [...], [...]])
converted_w_matrix = numpy.dot(matrix, array)
print(converted_w_matrix)
MrLixm commented

Fix released in 0.5.1
https://github.com/MrLixm/AgXc/releases/tag/v0.5.1

Thanks again for the report, that is a pretty embarrassing issue, would have been very annoying to not get this fixed.