[argyllcms] Re: xicclu -g predictability issue

  • From: Elena [service address] <1007140@xxxxxxxx>
  • To: argyllcms@xxxxxxxxxxxxx
  • Date: Tue, 25 Jan 2011 14:18:14 +0100

Hello Graeme

On 25-Jan-2011, Graeme Gill wrote:

> Sure. I agreed that adjusting K in the whole gamut to meet smoothly
> with the K at the gamut surface would be a next step, in an email
> some time back. That would help many cases, but doesn't address
> the issue of K on the gamut surface being "bumpy" itself.

What to tell... maybe "search for all matching candidates. If the first chosen
one generates a severe discontinuity, then chose the second one. And so on"...

> I don't think that the standard imposes any such limit, it's a matter of
> practicality. I doubt that you could invert that many grid points using
> the current code, in a reasonable amount of time. If "ultra" with
> a res. of 45 takes up to several hours, then a full resolution table
> would take about two weeks...

mmm... then something strange is happening here ?!
Yesterday I computed a device link profile with -G -qh (maybe redundant) -r100 
maximum allowed, -r101 start being refused) in a time that I didn't measure but 
is infinitely faster than just generating the original -qh (33) profile !
Miracles ? Or just I going mad ? :-D

Btw, it's interesing to notice how the results match the intended behavior.
See these softproofs (made with BPC as usual in order to keep the dark area
with artifacts well visible). I chosen as subject one RGB gradient which I
judge the most "bastard" in showing possible profile unevenesses:

 [this converted using the original test5.icm profile, LUT res=33]

 [this one using a device link profile created from srgb to test5.icm,
  with -n -qh -G -ip -cmt -dpp -k 0 0 1 .2 .5 which are the same parameters
  used for creating test5.icm, converted with cctiff. The only difference
  is substantially due to lack of conversion and interpolation in the
  intermediate Lab space]

 [same as above but using devicelink created with -r51]

 [same as above but -r100]

As you can notice, incrementing the grid resolution has the predicted effect
of reducing the bumps to thinner and thinner outlines - just because the
number of interpolated points around the offending knots progressively
diminishes. All right. So with a -r256 [currently refused by collink] I should
definitely see no artifacts at all!

So far... I don't like having discontinuities in my profiles, of course,
even because I could be wanting perhaps to convert 48 bit images at some point,
but as trade off I would be happy enough. Since you told you don't think the
standard imposes limits for the LUT resolution, and since [to be understood yet]
I could generate in few minutes a device link with 100^3 points, why don't
allow -r upto 256 ? Nice, huh ? :-)


Other related posts: