Comparing printer profiles neutral output

[rasworth]

I'm the proud owner of three printer profiles, for the same printer/paper combination, all generated by i1Profiler using the same spectro and parameter settings.  One is a year old, done with an i1Profiler 905 auto generated patch set, the other two were created recently using patch sets from Doug Gray, from his "... Optimal Profile Patch Sets" thread.

Test prints from the three profiles all appear identical to me, paying particular attention to the neutrals.  I came up with the following methodology to wring out some quantitative data without further printing:

1. Use a 33 step R=G=B image, from 0-0-0 to 255-255-255 in equal steps, in the ProPhoto color space.

2. Do an absolute convert in Photoshop for each of the three printer profiles, ending up with RGB triplets representing the raw data input to the printer for each step.

3. Arbitrarily assign the printer profile from the first set to the other two, retaining the RGB values and providing a means of visual comparison.

4. Paste all three images together.

Attached is the result, with the printer profile embedded in the image.  One can download and compare the actual profile outputs to the printer, as long as there is no conversion to another color space.  The colors of the patches have no significance, other than to provide a visual comparison of the RGB values.  In an ideal world I believe the outputs for each horizontal patch set should be equal, obviously they vary.  Without further printing and measurement I can't state which is the most accurate profile.  One conclusion is I'm probably not very good at judging test prints, particularly for the 376 patch profile there are some significant output differences.

The test prints were all done with relative rendering intent and black point compensation turned on - a question for the members is would this tend to hide output differences, or is the absolute output a valid comparison technique for visual results?

Richard Southworth

[Mark D Segal]

For a rigorous proofing exercise, where you are measuring the accuracy of printed values relative to file values, you should print with Absolute Rendering Intent. For a visual exercise where the objective is to see which solution "looks better", you would want to evaluate it using the same parameters you use for printing regular photos, so if that happens to be Relative Intent with BPC, you would use those settings for printing and viewing these patches.

If the objective is to visualize neutrality (or measure for neutrality), then of course colour matters, because the colour needs to be a*0 b*0 for the patch to look or print neutral. In your sample, the far right (376 patches) is less consistently neutral than the other two and the differences are visible, and confirmed running the Eyedropper down the strips.

[Doug Gray]

The profile attached to the images has numerous, bad patch reads in the 905 patch profile data. Lot's of black patches where there should be different colors.

Try this. Drag the profile (make sure you use exactly the same one you attached to the image) into the "Measurement" box in I1Profiler. It will populate the display with measured data just like you had finished scanning the targets. Look at the second page. If you make a profile from the data, the gamut plot it shows will have big holes punched in it. Printing any sort of normal color image should produce really awful prints.

Aside from that, the 905 patch profile's neutral axis shows more variation than I see from the Pro1000.

To get meaningful data from assigning a profile to an image in device space, you should use a gold standard profile otherwise you are just adding errors from both the source and destination profiles. Sometimes they cancel, but usually they magnify.

The tone curve seems way off. Convert the image back to ProPhoto and compare the rgb/Lab values. Way off.

Best way to evaluate a profile is to print the patches using Abs. Col. and read them with a I1Pro 2 using "Measure Chart" in I1Profiler. Save the Lab values and compare them to the Lab values you read out from the original ProPhoto image.

[rasworth]

I obviously did a poor job of explaining what I was trying to accomplish.

Mark,
I was trying to compare the absolute rendering output of the three profiles, the raw printer input rgb values for each patch.  The chart provides a quick visual check of equality across a horizontal set, without having to eyedrop for actual rgb values, otherwise the "colors" are meaningless.  Sorry, won't try that approach again.

Doug,
The profile works well in practice, without numerous black holes in the scanned target data.  Something must have been lost in translation.  I did as you suggested to check.  And I do understand that actual measurements are needed to show accuracy, I was attempting to find a means of comparing profile absolute neutral output variation, nothing more.

The original input image used in the absolute rendering conversion is attached.

Richard Southworth

[rasworth]

Doug,
Attached are the three profiles I used for the comparison.  Two of them were made with your cgats data.
Richard Southworth

[Doug Gray]

Doug,
Attached are the three profiles I used for the comparison.  Two of them were made with your cgats data.
Richard Southworth

Looking at the gamuts in PM5, the one from your 905 patch set matches reasonably closely with the two from my RGB sets. But there's something weird about the data in the profile. It's got a bunch of patches that are near black but shouldn't be. You should be able to replicate it by the following.

Open up I1Profiler. Drag the RAS_Pro100_CanonSemiGloss_Qual1_Cut_3-28-18.icc onto the "Measurement" tab. Go to the test chart and set the spacings the same as when you printed the target. Then go back to the measurement tab. Notice all the weird black readings.

The really odd thing is that the spectral data stored in the profile is what's corrupted. The profile appears to be fine. This would likely show up if you used the profile, since it contains the spectral data, to make a new profile with different settings. This is commonly done to try out the various perceptual intents, Contrast, Saturation, Neutralize gray. If you make a new profile from it it will be really bad!  Looks like some obscure bug in I1Profiler but I've never seen this occur and I've made a large number of profiles as well as used saved data in profiles to make new ones to test out the above.

Here's a screen capture of what I see. You should be able to duplicate what I see.

[rasworth]

Doug,

Mea culpa.  Not only are the invalid patches in the profile metadata, there also in my saved measurement data.  I counted 14 of them, scattered randomly thru the pages (see attachments).  I can only assume I had a bad day.  The profile was created a year earlier, couple of iProfiler versions ago, but I can't think of any mechanism whereby the invalid patches weren't included in the construction.  Perhaps i1Profiler is "smart" enough to discard?  And there is also the question as to why i1Profiler didn't flag the bad measurements during scanning.

I checked the other 20 measurement sets, including the two done with your cgats files, and didn't find another instance.  I'll use your 8/7 890 patch version for future printing.

Thanks for being vigilant,

Richard Southworth

[rasworth]

The mystery deepens, but I don't feel quite as incompetent.

I stuck the dongle into the pc and ran another profile with the suspect profile measurement data from a year ago.  Bad output, see attached ColorThink screenshot.  The stock profile does not look like this, even though it contains the invalid patches in the metadata.

So how can I have both bad measurement data and bad profile metadata and still have a viable profile?

Richard Southworth

[Doug Gray]

The mystery deepens, but I don't feel quite as incompetent.

I stuck the dongle into the pc and ran another profile with the suspect profile measurement data from a year ago.  Bad output, see attached ColorThink screenshot.  The stock profile does not look like this, even though it contains the invalid patches in the metadata.

So how can I have both bad measurement data and bad profile metadata and still have a viable profile?

Richard Southworth

I don't see any way this could be something you did Richard. Scanned patch values look fine except for those occasional bad patches and they don't look like operator error or instrumentation issues. They smell like I1Profiler software bugs. The fact that the original profile does not reflect the saved spectral data is a very strong indicator of software problems.

I'm starting to appreciate my workflow more. I scan targets in both forward order and reverse as well. The charts I print have registration bars on both ends of the paper so it can be fed in from the top or bottom. So, for the second pass, I feed the paper in flipped around and go from the last page to the first. This results in a second CGATs file that is, if everything is working, exactly the same except everything is reversed. Then my Matlab script just reverses the latter and averages as well as compares the results for consistency.

[Doug Gray]

I find this image handy for checking neutral tones with a handheld spectro. It's in 16 bit ProPhoto RGB and it has L* steps 0,1,2,3...99. Each row starts at increments of 10. It's in 10mm squares, high precision and the L* values are accurate to well under 0.1, typically less than +/- 0.02. Photoshop throws away the LSB of 16 bit images.

Print it using Absolute Colorimetric. It's best to set the color settings: Edit->Settings->Use Dither otherwise photoshop will round the RGB values to 8 bits when using Windows. However, even with 8 bits the rounding error is small and a* and b*, which are more important perceptually are not affected.

[Alan Goldhammer]

I find this image handy for checking neutral tones with a handheld spectro. It's in 16 bit ProPhoto RGB and it has L* steps 0,1,2,3...99. Each row starts at increments of 10. It's in 10mm squares, high precision and the L* values are accurate to well under 0.1, typically less than +/- 0.02. Photoshop throws away the LSB of 16 bit images.

I'm only using Argyll freeware to do things.  I can print out a 100 step B/W patch set using appropriate commands.  I just checked RGB values in Photoshop for some random patches and they look OK.  Can such a test chart be used in the same manner as you describe for iProfiler?

[Doug Gray]

I'm only using Argyll freeware to do things.  I can print out a 100 step B/W patch set using appropriate commands.  I just checked RGB values in Photoshop for some random patches and they look OK.  Can such a test chart be used in the same manner as you describe for iProfiler?

For neutral tracking, I use a patch set of 95 neutrals, repeated 9 times then randomized. For the highest precision, I've found it necessary to randomize patches as the values have some degree of printed location sensitivity. Repeating each patch 9 times provides much improved statistics, and reduces the standard deviation of the mean by about a factor of 3.

This was all well below anything I could see. At the time I was making specialized charts and was focusing on repeatability.

Unfortunately, while such a chart is great for fine tuning where I needed to add additional, fine grain near neutrals to a target, it is not useful for a visual check.

The 10x10 neutral Lab image is a good, fast, visual check and it can be easily measured with a hand spectro. One can do the whole set or just measure every 5th or 10th value and plot it for faster results. I don't use it that often but it's great for things I don't have a profile for and/or the iSis can't scan like really thick media or canvas..

[rasworth]

Doug,

Downloaded the test chart, printed absolute using profile created with your 8/7 patch set, and drying.

Couple of questions,

1. Is it practical to include patch separation lines?  Hard to distinguish individual patches, particularly in the darker tones.

2. Since I'm lazy, how about in scannable form, patches sized for i1Pro?  Who says beggars can't be choosers.

Thanks again for your help,

Richard Southworth

Added by edit - Just read your last post, I assume the separation lines would mess up the visual check, never mind.

[Stephen Ray]

the iSis can't scan like really thick media or canvas.

I know those who use the iSis to scan canvas often. Can you elaborate?

[Doug Gray]

I know those who use the iSis to scan canvas often. Can you elaborate?

The only canvas I use is rolls on the 9800. I briefly tried running iSis targets for it but found significant errors. Crosstalk from neighboring patches. I use a chart with registration bars on the top and bottom so I can feed the paper in bottom first to make a second data set then compare them for errors. Chalked it up to registration issues due to its stretchiness. Since I wasn't using canvas for high accuracy technical chart work, and found the Epson canned profiles quite good, I didn't pursue it. However, I only tried the 6mm patch size.

[Doug Gray]

Doug,

Downloaded the test chart, printed absolute using profile created with your 8/7 patch set, and drying.

Couple of questions,

1. Is it practical to include patch separation lines?  Hard to distinguish individual patches, particularly in the darker tones.

2. Since I'm lazy, how about in scannable form, patches sized for i1Pro?  Who says beggars can't be choosers.

Thanks again for your help,

Richard Southworth

Added by edit - Just read your last post, I assume the separation lines would mess up the visual check, never mind.

I don't have such charts Richard. I've found it necessary to scramble patches for the I1Pro as it seemed to have a hard time with small changes in adjacent patches. Since I prefer scrambled patches and have the software such that descrambling is automatic, I really haven't pursued it. Been a long time though and it may well be possible with the I1Pro2 and current software.

[rasworth]

Test print (dye inks, Pro 100, Canon semi-gloss) probably hasn't dried long enough for accuracy, but I decided to run a set.  Again the test was using an i1Profiler profile based on Doug's 8/7 890 patch set.

I1Pro uv cut, spot measurements, down the +4 column:
 
L*in            L*        a*         b*

 4            5.35      0.17     -1.67
14          13.57     0.15     -0.55
24          23.02     0.12     -0.41
34          33.00     0.66     -0.45
44          43.69     0.51     -0.33
54          53.94     0.18      0.06
64          64.01     0.09     -0.01
74          74.27     0.29      0.26
84          84.91     0.29     -0.02
94          94.75    -0.07     -0.70

So what is "goodness"?  Visually I don't see any tint.

Richard Southworth

Added by edit - Sheet of same paper under test print during spot measurements.

[Mark D Segal]

Test print (dye inks, Pro 100, Canon semi-gloss) probably hasn't dried long enough for accuracy, but I decided to run a set.  Again the test was using an i1Profiler profile based on Doug's 8/7 890 patch set.

I1Pro uv cut, spot measurements, down the +4 column:
 
L*in            L*        a*         b*

 4            5.35      0.17     -1.67
14          13.57     0.15     -0.55
24          23.02     0.12     -0.41
34          33.00     0.66     -0.45
44          43.69     0.51     -0.33
54          53.94     0.18      0.06
64          64.01     0.09     -0.01
74          74.27     0.29      0.26
84          84.91     0.29     -0.02
94          94.75    -0.07     -0.70

So what is "goodness"?  Visually I don't see any tint.

Richard Southworth

Added by edit - Sheet of same paper under test print during spot measurements.

From the point of view of "tint" goodness is those numbers you report. This is really good performance. One wouldn't expect to see a tint based on that data.

[rasworth]

Mark,

Thanks for the input, good to know.

Someday I'll run the 10/9 1764 patch set, 5 letter size sheets with an i1Pro is no fun.  Is there anything between the i1Pro and i1iSis to facilitate this process?  I'd love to have one of the latter, but there's no way I can justify, and its purchase would probably initiate divorce proceedings.

Richard Southworth

[Doug Gray]

From the point of view of "tint" goodness is those numbers you report. This is really good performance. One wouldn't expect to see a tint based on that data.

Yep, you won't see tint with those numbers. The only place a* and b* get  slightly high are at L=4 and 94. This is normal because you are near the black point and white point. You will get increasing deviations at lower and higher L* but mostly in the L* itself.  Normal and unavoidable. Rel Col doesn't have this at the high end and Re. Col. with BPC doesn't have this across the full range since these track against the paper's white point. IT is, however, harder to get a good metric because spot readings don't directly correlate. You have to scale the readings against paper white.

The numbers are consistent with what I see on the Pro1000 as well.