Re: [Lcms-user] Creating equivalent matrix-shaper and 3D LUT profiles, problem with BPC

[Pekka P. <pek...@ha...>]

On Sat, 16 Mar 2024 10:09:29 +0100
<mar...@li...> wrote:

> Hello,
> 
> Well, no solution fits everything, of course, but I found the CLUT plus
> linearization curves to work fine in many scenarios. I think the issue here
> is HDR is not well supported by ICC workflows, highlights and values above
> L* 100 of D50 are seldom discussed in the 4.4 ICC spec. For BT.2020/PQ,
> which profile I'm attaching this is quite evident. Lcms can deal with it on
> unbounded mode, but if you feed a RGB -> Lab transform, for
> RGB=(255,255,255) you will find L*=523!! This is not encodable in any CLUT
> unless you provide enough headroom. 
> 
> I did this paper many, many time ago regarding how to increase accuracy of
> CLUT. Hope that would be of any help.
> https://www.littlecms.com/ASICprelinerization_CGIV08.pdf

Hi Marti,

interesting indeed. The idea seems to be to separate the per-channel
"average" approximate non-linearity from the rest of the
transformation, assuming that the uniform distribution of the sampling
points of a 3D LUT is then better suited for the rest of the
transformation. When I think of the kind of color transformation
pipelines from source space to blending space, they do seem to be
exactly that form until tone and gamut mapping.

That does seem worth studying in the context of HDR transformations to
evaluate achievable precision. I would be surprised if someone hadn't
already done that, but I don't know where to look.

Weston should probably apply that method in its 3D LUT path anyway.


Thanks,
pq