Elena [service address] wrote:
I know it may sound a blasphemy and it will hurt purists... the trick is simply using an arbitrarily high level for -r in colprof, eg. -r5. I know the major side effect here is having a generated A2B table which won't follow so precisely the device response anymore, and I could hate that... but I tried such a profile for my plain paper tests made with the "offending" -kx and -r5: I don't see the bumps anymore ! And xicclu now shows a K which is monotonically increasing.
It's quite hard to algorithmically decide the level of smoothing that optimizes overall accuracy. I've had a go at it many times, using several approaches, with several results. It certainly seems that on some occasions the level of smoothing is underestimated (some XYZ based cLUT profiles come to mind), although I think that you are probably going to extremes. I can imagine that doing so does to some degree overcome black topology problems, and smooth out the K plane in the result as well, but at a terrible cost to colorimetric accuracy.
So I now wonder (since other people might want to follow this approach more or less at their own "risk") if it could make sense for you to "split" the -r behaviour independently for A2B and B2A tables generarion. So one could still chose precise A2Bx tables (eg -r1 =0.5) and artificially smoothed B2Ax ones (eg -r2 =4 or 5).
Not really. There is only one A2B table computed, and it's the smoothing of that, that you are changing. All the different B2A tables are created from that one A2B table. Graeme Gill.