Ben Goren wrote:
Well, as I mentioned, I was able to get a profile with an average error of ~0.8 using an XYZ LUT.
In my experience (with scanner profiles), the low self-fit error of XYZ CLUT profiles can be deceptive.
Although the self fit error of Lab CLUT profiles is higher, they can perform better than corresponding XYZ profiles when one uses independent test- and trainingset (see example below):
a) testset == trainingset: profcheck -k TI3/combined.ti3 ICC/combined_Lab.icc Profile check complete, errors(CIEDE2000): max. = 9.178650, avg. = 1.371444, RMS = 1.841394 profcheck -k TI3/combined.ti3 ICC/combined_XYZ.icc Profile check complete, errors(CIEDE2000): max. = 8.111750, avg. = 0.939068, RMS = 1.496107 b) testset <> trainingset: profcheck -k TI3/testset.ti3 ICC/trainingset_Lab.icc Profile check complete, errors(CIEDE2000): max. = 9.018731, avg. = 2.481812, RMS = 3.028049 profcheck -k TI3/testset.ti3 ICC/trainingset_XYZ.icc Profile check complete, errors(CIEDE2000): max. = 21.973729, avg. = 2.892268, RMS = 4.078021 Klaus