Hi, Nikolay Pokhilchenko wrote:
Yes. The printer is network one. I have no rights to change it's settings.
I disagree with you for this certain case: when the initial target was not optimized across perceptual space, the really-problem patches may have a little preliminary-profile fit error in the device space. This approach will work only for the space the target optimized is. With your recommendations I'll be able to see errors in the device space, but it didn't mean I'll see errors in the color (perceptual) space. I think when the device is quite nonlinear, the quality evaluation may be done only in perceptual space, nor the device space.
You may be right there - my reasoning was just that the driver has a built-in mapping between the device space and output space. This mapping is fairly simple in places and pretty complicated in others, and since you can't disable it or select a simpler mapping, your profile needs to track its behaviour as well as the printer's behaviour.
I agree that in general the fit errors don't tell you much about absolute accuracy - but I think they do tell you where the model is having particular trouble fitting the data points, and my assumption was that the profile's model will have the most trouble fitting data points in the "complicated" areas, where there will be some benefit from concentrating extra patches.
For given target data it improves the device behavior characterization, but will be less useful for resulting color images.
Because smoothness may be compromised? I agree that it's certainly possible to over-fit a model, but when you have a worst-case dE of 8.9 it's probably safe to say you're not over-fitting yet!
Incidentally, I don't think you mentioned the actual problem with the profile you generated, beyond it being insufficiently sampled in certain areas. If you actually print with it, what's wrong with the results?
Alastair, what aim do you achieve such way? The good developing of a grayscale?
Yes, mainly - and also adding some near-skin-tones in situations where a profile is required for giving "nice" rather than "accurate" output in situations where it's impractical to take "enough" samples. (Inkjet-printable CDs, for instance.)
I wrote some month ago about my way: the cylinder, randomly filled by patches around the gray Lab (or Jab) axis.
Oh yes, I remember reading that now.
I think the best profiling way is to print both the device space and the perceptual space optimized tagets. IMHO this approach can improve both the device and the perceptual behaviors - both the A2B and B2A tables.
Yes, it makes sense - is that not in essence what the ColorMunki's profiling software does, with its two-pass system?
All the best, -- Alastair M. Robinson