One more final thought, that the final thoughts below brought up...
There is a difference between accuracy and precision. In a lot of cases 7 or 8
bit accuracy is enough because you are not comparing two colors side-by-side
only that they look the same as you intended. Prints and displays also require
different system gammas (typically 1 vs 1.5 respectively) to look the same
based on several studies so there is also non-linear processing that has to
happen in the driver and how people perceive the match between monitor and
display also varies so worrying too much about accuracy is counterproductive.
On the other hand, to prevent quantization artifacts, you may need between 8 to
over 20 bits of precision depending on the color space transformations required
and display conditions. Under the right conditions in digital cinema you need
12 bits of precision for quantization artifacts to never be noticeable under
any condition. However for prints, 8-9 bits will always be enough because there
is always a reference white in the field of view that increases the just
noticeable difference that you can detect.
Technically, 8 would be enough for printing if it wasn’t for the change in
system gamma. Because of that, if you start with 8 bits of precision, then some
areas will end up with 9 or more and some areas will have 7 or less.
My point being that even if you start with an 8 bit image, I would keep all the
precision of the 16 bit file format and make sure you have a 16 bit printer
driver so that you never have to worry whether or not the changes made in
editing or printer cause quantization artifacts.
On Oct 16, 2018, at 6:18 AM, Jason Campbell <campbell.jj76@xxxxxxxxx> wrote:
Yves/Graeme,
Yes I was also quite interested in that detailed writeup. It made sense, but
is impractical for all the reasons everyone stated. Most notably that even a
high end spectro has a repeatability in the 0.01-0.05 dE range. Combine that
with just the nuances of measurement and there’ll be a good bit of noise in
the data. But Graeme’s point made sense -- you have to have a lot of samples
to combat that.
Further, I’d agree in general purpose printing, it might be pretty hard to
get something to exhibit the benefit of a 16bit pathway. The only reason I
was looking into it was that when you do profile lookups, you can get
fractional data out (e.g. Lab->RGB). So my hope was that instead of clamping
values to an 8bit integer, I could pass through the interpolated value from
the profile in a way the driver could “do better”. Meaning if I was bouncing
around 0.8-1.0 dE with integers, that getting higher-resolution RGB values
into the driver would let me eek out another little bit of dE.
Though I’ll say in some testing, that just the general ‘bounce’ of print to
print around a large sheet might vary 1.0 dE depending on several variables.
But then again all the more reason I was on the hunt for getting as close as
possible via a 16bit path.
From: "argyllcms-bounce@xxxxxxxxxxxxx" <argyllcms-bounce@xxxxxxxxxxxxx> on
behalf of Yves Gauvreau <gauvreau-yves@xxxxxxxxxxx>
Reply-To: "argyllcms@xxxxxxxxxxxxx" <argyllcms@xxxxxxxxxxxxx>
Date: Tuesday, October 16, 2018 at 8:50 AM
To: "argyllcms@xxxxxxxxxxxxx" <argyllcms@xxxxxxxxxxxxx>
Subject: [argyllcms] Re: testing
Graeme,
I kind of wrote this as a thank you all for all your replies and as the show
must go on kind of thought.
But what you wrote here intrigues me in many ways and your approach
basically close the subject on this 16 bit printing stuff. Using your metric
one would need an instrument capable of measuring Delta E of well below
0.005 units. The one I intend to buy can only go down to 0.1 under the best
condition. In other words 16 bit printing is most likely just marketing BS.
But I assume if not really 16 bit, maybe 9, possibly 10 would still be an
improvement. My goal is still the same, I shoot a lot of flowers and
critters of all sort and I would like to print as much as possible the
colors and or gamut that I captured with my 14 bit sensor. I'm only looking
to get the most of these images. Before you mention it, I know perfectly
well my eyes (I'm 60+), my monitor and probably my printer as well can't
print much if anything above 8 bit anyway? So I'll concentrate my effort
more on refining as much as possible the profile I use for printing and
finding better papers. I read a post elsewhere that gave a hint on how to
reduce if not eliminate 8 bit artifacts like banding. I'll be on the lookout
for anything that can help.
Many thanks Graeme, your approach closed the subject on this matter for me.
Thanks again,
Yves
-----Original Message-----
From: argyllcms-bounce@xxxxxxxxxxxxx [mailto:argyllcms-bounce@xxxxxxxxxxxxx]
On Behalf Of Graeme Gill
Sent: Monday, October 15, 2018 7:57 PM
To: argyllcms@xxxxxxxxxxxxx
Subject: [argyllcms] Re: testing
Yves Gauvreau wrote:
At first, I kind of wondered if in a forensic kind of way it was possible
to distinguish
between a 8 and a 16 bit print. It seem, it could be a waste of time to
even try. But I'll
still get myself an I1 Studio and try my best to get the most out of my
work at least
color wise.
should be possible, but it depends on how much trouble you are prepared
to go to. i.e., rough outline:
If you assume that the gamut of a printer roughly spans 150 delta E units,
then 8 bit precision amounts to 0.6 delta E if the space was perceptually
uniform. So you might aim for a measurement precision of 0.2 delta E
or better, so as to be able to distinguish 0.6 delta E steps.
Say you pick a test color which you guess will be sensitive to
quantization, and print and measure it 10 times, and then
calculate the standard deviation as 0.9 delta E. To
get a standard error of 0.2 you will need to measure
and average (0.9/0.2)^2 = 20 or more samples.
So then you would print a series of test patches,
(say) 10 that span two 8 bit precision values,
print and measure each 20 times (total 200
patches to print and measure on 10 sheets) and plot the
average of each 20. If the graph is continuous,
you can assume that the printing system has better
than 8 bit resolution. If the graph has one distinct step
in it, then the printing system has 8 bit resolution.
Some of the quandaries are that the space may not be
very uniform, and you really want to test a color
that is most likely to be quantized than not.
You also can't really be certain if a quantization
step you have found is in the the input to the
printing system or is generated within the
printing system itself in the process of screening etc.
Cheers,
Graeme Gill.
