[argyllcms] Re: Verifying profile quality of LUT-based scanner and printer profiles

Milton Taylor wrote: 
The more you learn the more questions it raises...!

OK I think I'm getting closer to understanding what's happening here. Certainly the banding and posterisation that is happening on the Hue/Sat rainbow has a lot to do with gamut clipping. I can see that this is indeed a savage test, but a fair amount of colours [mainly the greens and blues] even down to Saturation=50% are out of gamut. (At L=0.5) I don't believe this is peculiar to this printer though.

But this has exposed what I think may be the real underlying problem: the 'smoothness' of the contours that define the device's gamut. It seems to me that at the moment, the posterisation and banding I am seeing in the rainbow could be caused by sharpish 'angles' in the contours of the gamut's skin (thinking in a 3D sense), leading to weird effects by the gamut clipping routines when rel col is used. I gather this is what you referred to as something you were working on? 

There are lots of things that affect the result. The precision of the
gamut surface determination. The smoothing factors used in profile
creation (the gamut of a profile with too lower smoothing is pretty
obvious when you learn what to look for. Most profiles should
resemble something like sRGB in terms of the crispness of the edges,
and the smoothness of the surfaces. If the edges look ragged, or
the surfaces have big "dings", then it could be that the smoothness
is too low.) The way the mapping guide vectors are created, and how
they are smoothed. The smoothing factor of the spline gamut mapping
function  creating from the guide vectors. The precision loss in
using the B2A table of a profile, rather than creating a device
link by direct inversion.

The smoothest result is achieved on the sRGB gamut being printed as Perceptual rendering, although some posterisation is still evident. Also, the color range is now woeful... all the deep blues and greens have gone! Clearly the shape of this printer's gamut (or what the profile think's is the gamut) must be cigar shaped, and perceptual is sacrificing one color dimension far more than the other in order to make the whole thing fit, and still preserve color relationships. 

Generally additive and subtractive gamuts are opposites. They have
complimentary primaries, and secondaries are usually less saturated
than primaries. Additive primaries are generally lighter than
subtractive primaries, and subtractive secondaries (the colors
that correspond to the additive primaries), are even darker. Add
into the mix the manufacturers mapping of RGB to CMYK, and
there are a lot of tradeoffs.
Possibly you're more interested in saturation that lightness matching,
so the current releases emphasis on lightness matching doesn't suite
your goals.

I was a bit surprised to see that a fair swag of the color space in sRGB falls outside the printer's gamut. For some reason I had assumed that sRGB fitted reasonably well into the typical inkjet's gamut, but that is apparently not the case. This is where I need to look at the 3D shape of the gamut compared to other printers and to the various working spaces. 

It's hard to simplify things like saying what's a big gamut or not. Most
additive device gamuts are big, and printer gamuts are often more limited.
This is basic physics. Additive devices can make their primaries as narrow
band as they like (they can use lasers if they like), meaning that their
primaries can lie extremely close to the spectrum locus, and enclose a lot
of space. Subtractive devices can't make their primaries narrow, or the
result would be much too dark. The best they can do is make them as
"block like" as possible. The way the subtractive primaries mix
causes the secondary colors to not usually lie within the triangle
formed by their primaries, the way that additive devices do.

The opposite nature of the two types of device primaries, and
the different nature of how the primaries mix, usually means that
additive and subtractive devices gamuts each have areas not covered
by the other. The claim that "sRGB has a really small gamut" is
rather a simplification and exaggeration. What perhaps is really
meant, is that it does not cover a lot of gamut that is captured
by photographic film, or printable by many printers. There is still
a lot of sRGB gamut that can't be printed on any printer!

If the shape of the gamut skin has any significant non-linearities in it, I imagine it is possible that these are not necessarily smoothed out just because of using Perceptual rendering intent? Or, is it possible that you can still get banding with Perceptual rendering intent for some other reason? 

Perceptual intent at least tries to smoothly bend the source gamut
to fit the destination gamut. Colorimetric simply maps to the
closes destination color, clipping if it is out of gamut. There
will therefore be an edge when clipping starts, and (worse than
this), the shape of the gamut may be such that there are abrupt
transitions from one point on the destination being closest,
and another point being closest (think of a Voronoi diagram).
There is no attempt to smooth such effects for the colorimetric intents.

Another question: all this was done at 8 bits per channel. I am wondering if I started out with a 16 bit rainbow, would I possibly get less errors being introduced when Photoshop does the conversion to output space, even though the printer will only be taking it at 8 bits? I suppose that is possible when working at the extremes of the gamut where the RGB values are either almost zero or almost 255 on one of the channels? 

Possibly, but generally banding effects due to 8/16 bit are relatively
subtle, and more often show up in rendering neutral ramps etc.

Lastly, I wonder if this is a case where perceptually, the eye is more concerned about the posterisation bands than it is with the spatial size of the colour flat spots caused by clipping 

Certainly the edge sensitivity of the eye (mach banding) draws notice to
such defects.

> Perversely, a profile with
smaller numbers of entries in the LUT may actually cause less noticable banding than a high quality LUT. Mind you, if the objective is to always stay inside the printer gamut then this is a non-issue. 

This is often the case. For most uses, I'd tend to use "medium" quality,
and only try something else if there is a particular reason to do so.

Graeme Gill.

Other related posts: