While RGC works really nice, I feel like it’s still solving the problem in the wrong domain. And while a lot of the artifacts in out-of-gamut gradients can easily be solved before IDT+RGC, it’s almost impossible to fix them with the regular color grading tools after.
I have no idea what it could be, but 3x3 matrix feels a bit …old maybe?
Intuitively I see the new approach to be like a 3x3 matrix for regular colors, but at some point somehow interpolating to meaningless highly saturated colors of the source gamut. They can be too off. But if this could solve the current issues, I think, shifting them to somewhat more pleasant is a way more easy task.
And here is another proposal:
To add parametric RGB sliders into IDT that work in linear gamma, but in source camera primaries. While they are not identical to the actual sensor primaries, I think (or hope?) It’s still closer than AP1 or any other unified color space.
The ability to use some operations before IDT is the reason (apart another Resolve bugs) why I build ACES using nodes instead of using it as intended (selecting ACES in color science in Resolve project settings).
Hi Anton,
There are plans for better IDTs, things like 2.5D LUTs or Root-Polynomial Colour Correction are things on the radar.
The 3x3 matrix offers a lot of advantages: It is fast, it is trivially invertible and it is exposure invariant. Those are 3 critical aspects to consider.
Cheers,
Thomas
The 3x3 matrix not working is a symptom of overall issues and not a cause.
The 3x3 matrix is classically used in trichromatic color systems because it is a clear way to transform between the coordinates of linearly related color systems.
The less the linear relationship between two color systems, the less applicable the solution will be. It can still work at points, but the overall solution will break.
In general designing a system where the all tonality points between the two systems is positive is very helpful. Since the system described by a 3x3 matrix multiplicative, negative linearization values are not describable.
So simplify the problem think about a 1-d space transformation where the source to destination transform is one stop different dst=2*src
src=1 dst will be 2
Now of src =-0.01 dst will be -0.02
The implication of this is that the src can’t be realized in the destination space.
The way to work around the first issue in ACES is to transform the spaces so that they are all positive , then at the end scale appropriately. If a camera log space ever results in a negative value though an idt add values in the camera log space until the smallest value is positive, after the transformation to linear use a multiply to reduce the exposure back to the original.
Basically things will work much better switching from a relative linear solution, as exists now, to an absolute linear.
The reminder of the problems are more general to color management and require, bespoke, multichannel or spectral solutions.