[argyllcms] Re: How lumpy is the LUT produced by colprof?
- From: Graeme Gill <graeme@xxxxxxxxxxxxx>
- To: "argyllcms@xxxxxxxxxxxxx >> argyllcms" <argyllcms@xxxxxxxxxxxxx>
- Date: Mon, 4 Sep 2017 21:34:08 +0200
Jarkko Lavinen wrote:
Hi,
There is a smoothing option -r (and -rr, rs) in colprof but ti3 file
does not seem to carry individual patch XYZ noise stats and the
smoothing figure likely does not depend on how reliable the patch XYZ
reference data is. Could this exaggerate bumps with noisy patch XYZ
reference data?
Maybe - but you are jumping to conclusions. 98% of the time "lumpiness"
is due to some issue with poor device setup or details of how the patches
are being produced or measured.
I was initially using Vuescan with IT8 calibration and scanning slides
and thought everything was fine. Then I learned Vuescan supports only
its own matrix profiles and they suck and LUT profiles are supposedly so
much better [1] [2]. And then finally I learned LUT profiles suck too
[3] [4].
Nothing's that black and white. Each approach has trade-offs.
Oldendorf is the author of ColorPerfect PhotoShop plugin. He talks about
color integrity being fragile and being destroyed by non-linear
transformations. In [3] he says: "Lookup tables are completely
unsuitable for RGB input devices.". In [4] he criticizes many thins but
LUTs are criticized for their lumpiness and breaking havoc. Is this FUD?
I can't agree with such an extreme view. Shaper profiles generally provide
smoother behavior, but less precision and detail. cLUT profiles generally
provide higher precision and more detail, but (at least for ICC style
standards) are piecewise linear, so are technically not smooth.
[ But remember that any shaper profile that is fitting a real world
device is likely using 1D piecewise linear lookup tables as well! ]
The practical point is not that lookup tables are technically
not smooth, it is whether they are practically not smooth. Any
slope discontinuity will be invisible below a certain threshold.
(And I'm referring to ICC style lookup tables here because it's
certainly possible to use lookup tables in ways that ensure
2nd order continuity, but they are slower.)
But it seems that you are particularly referring to input profiles
rather than color profiles in general, and most readily available
input charts are not very suitable for creating cLUT profiles, simply
because the number and arrangement of patches is too sparse, It8
charts being especially bad. Some people have gone to the trouble of
creating their own input charts with a much larger number and more
even spread of patches, and smoothness is much less likely to be an
issue, regardless of choice of Shaper or cLUT profile model.
Cheers,
Graeme Gill.
Other related posts:
