Relative vs Perceptual compared

[AndyF]

Everyone here quite likely knows the differences between Relative and Perceptual intents, but I created a different test illustrating the differences.  A short version of the test is, make RGB bars with increasing intensity in PS, then print first with Relative then (further down the same paper) with Perceptual intent.  Dry, then scan the print back in.  Custom software then samples down the length of the bars and plots the measured colours.

Examining the two charts, it can be seen that Relative tries to compress the entire AdobeRGB 0-255 range to preserve relative differences between all tones while Perceptual seems to try to keep the AdobeRGB values as much as possible in the lower ranges, until it approaches the maximum of the destination space, and then plateaus for the remainder of the tones.

The longer description is at https://andyfraserca.wordpress.com/?page_id=798

Andy

[digitaldog]

There are no rules in creating a perceptual rendering and every manufacturer does it differently. On top of all the possible settings when creating a profile with that RI. So YMMV.
Lastly, profiles used with any RI know nothing about images and color in context. You the image creator do while soft proofing and selecting an RI you prefer visually.

[digitaldog]

If you used either Adobes profiles (Adobe RGB (1998) or sRGB), you never got a Perceptual RI. There is no such table in either matrix profile. You can select Perceptual in PS indeed. You get RelCol.

There are many test image online; for this image I made a collage from parts of various colourful photos and added the Outback test image.  To generate the Relative and Perceptual versions of the image, the image was converted to the sRGB colour space in Photoshop using those two intents.

The three test images are shown below with the AdobeRGB version first, Relative sRGB second and Perceptual sRGB last.  All three will have now undergone another conversion into your browser’s colour space but it would be applied equally to all three.

[AndyF]

I just checked and you're right.  Using the "convert to profile", from Adobe RGB to sRGB, there's a choice of the Adobe ACE or Microsoft ICM engine, and Rel or Perceptual.  I would have assumed (was expecting!) Rel and Prcpt to be different and maybe Adobe to be slightly different than MSFT.  All four choices give exactly the same chart.  And none of them do a good job with red either... 

It's really annoying that Adobe would leave something non-functional in PS instead of just commenting out that bit of code to make the feature vanish until fixed.  Thanks for pointing that out Andrew.  I'll fix my last examples in my article over on Wordpress.

The attached chart is for all four of those cases.

[digitaldog]

I just checked and you're right.  Using the "convert to profile", from Adobe RGB to sRGB, there's a choice of the Adobe ACE or Microsoft ICM engine, and Rel or Perceptual.  I would have assumed (was expecting!) Rel and Prcpt to be different and maybe Adobe to be slightly different than MSFT. 

They can't, again, there is only one table (Colorimetric). Been this way since day one in Photoshop (5.0).
There is at least one sRGB profile (V4) with a perceptual table but this is all pointless anyway. Again, there are no rules in how anyone creates such a table. You could scan 10,000 color patches on a Spectrophotometer from a printer and feed that data to two different software packages to build a profile, even attempt the same settings for Perceptual rendering IF they existed, the two would not behave the same. Just as you can load Velvia and Ektachrome in two identical cameras, pointing at the same scene or target and get two different renderings.
Your 'testing' doesn't tell us anything that useful, other than indeed, there are differences in RelCol and Perceptual with the same profile. So what? One RI may produce a more pleasing conversion of one image while the other RI may do the opposite. Profiles know nothing about images. They "see" everything one pixel at a time. You the image creator see an image. You can and should soft proof, even examining the Saturation RI and pick one based on the image in context and what you prefer visually.
This has nothing to do with color accuracy despite your graph. This isn't about color appearance, this is about color perception of which all this is based on a single color under a defined condition.

It's really annoying that Adobe would leave something non-functional in PS instead of just commenting out that bit of code to make the feature vanish until fixed.

There is nothing to really fix; the profile doesn't have the table. I suppose Adobe could try examining each kind of matrix profile, then gray out Perceptual. But it doesn't matter really; pick Perceptual or RelCol, you get Relcol and that is what is soft proofed visually of course.
Bottom line: no rule in what RI to pick, view the conversions and pick what you want. Perceptual is supposed (repeat supposed) to render to perceptually deal (gamut compression) with out of gamut colors instead of gamut clipping them to the gamut boundaries and everyone does this differently. A better visual than a graph (after the reader excerpts there are no right or wrong answers here, it is subjective):

[AndyF]

I think the end result is, get one good custom profile for each printer/paper combination so the printer & paper are being used to it's fullest and most accurate capability in that part of the printing path, choose which RI works best for the image and know generally what mapping each RI applies, and make the rest of the adjustments intentionally in LR/PS.

[digitaldog]

I think the end result is, get one good custom profile for each printer/paper combination so the printer & paper are being used to it's fullest and most accurate capability in that part of the printing path, choose which RI works best for the image and know generally what mapping each RI applies, and make the rest of the adjustments intentionally in LR/PS.

Accurate capability is a big rabbit hole without actual data.

What we have here are three very different attributes of color: accurate color, matching color, and pleasing color.

Let's start with accurate color since you originally brought it up. Like gauging anything for accuracy; speed, distance, height, weight, we need to measure something and compare it to a reference to know if it's accurate. This is true for color. Let's say you show me two images on-screen of a red apple. The color of each apple is noticeably different. You ask me "which is the accurate color of the apple?" I can guess but I cannot answer it accurately without measuring the apple with an instrument designed for that task. A Spectrophotometer will provide device-independent numbers of the color, a value in Lab. I then need to measure the exact same spot** on the display for both apples and produce that Lab value for each. Now I can use a formula (they differ too) to calculate the distance/difference in the actual apple and the two images of an apple. If the value is greater than 1.0, it is said to exceed JND (Just Noticeable Difference). IOW, if the image on the screen and the apple are a detlaE (dE) of 1 or less, they appear the same to the observer. This isn't going to happen but the idea is, 1.0 or greater is a visible difference, higher, more difference. Less accurate between measured color and resulting color.

Moving to another kind of accuracy; measurement of distance. Let's say I measure my foot with a ruler and find it's 10.7 inches. Clearly, my foot isn't an adequate measuring device compared to a 1-foot wooden ruler. But how accurate is the ruler? It too should be of sufficient accuracy for the task. There are devices like a 50 cent wooden ruler that may be accurate to +/- 1/10 of an inch. For measuring a 40-foot wall, that may be all that's needed. For measuring a component that will be used on a spaceship, 1/10000 of an inch may just be adequate. But my foot clearly isn't. So when we talk accuracy, we need to take into account the instrument or method of measuring the accuracy. And how accurate we need the results of the measurement to the reference. With dE, 1 or less is invisible. There's a difference but for viewing a solid color, it's moot.

Color matching may have nothing to do with accurate color due to metamerism (which is good and provides a visual match) and metameric failure where we don't get a match but might have the ideal color 'accurate' values.
Pleasing Color is subjective and has little to nothing to do with any of the above. Back to Color Appearance vs. Perception.

**Getting back to color in context, viewing an entire object, print, display: An apple isn't a single color! So I'd have to measure many spots on the apple as well as the screen. Digital images are composed of thousands or tens of thousands/millions of solid colors. As would our apple. Zoom out of the image on-screen, it appears like an apple. Zoom in, it's just a big pile of solid color squares with color numbers. As you can see, discussing the accuracy of a photo of an apple and the apple itself is a big job!

[Doug Gray]

Color matching may have nothing to do with accurate color due to metamerism (which is good and provides a visual match) and metameric failure where we don't get a match but might have the ideal color 'accurate' values.

Small quibble. Metameric failure can, and too often does, occur with two colors that have been measured to have the same L*a*b* value, but appear differently when viewed under some LED or fluorescent light. However, they should appear the same when viewed with D50 and with the same surround.  So "Accurate" or even better "Consistent," requires good instruments and control over the lighting and viewing environment.

Because only Relative or the rarely used Absolute Intent are actually specified by the ICC, they are the best way to get the most consistent prints when printing the same image with different printers. And only if their profiles are reasonably decent and the image is in gamut. Adjust images in Photoshop or LR for the most appealing look soft proofing with RI.

[fdisilvestro]

I just checked and you're right.  Using the "convert to profile", from Adobe RGB to sRGB, there's a choice of the Adobe ACE or Microsoft ICM engine, ...

I would stay away from the Microsoft ICM engine. For instance it can show out of gamut colous converting from AdobeRGB to ProPhotoRGB because it does not perform chromatic adaptation related to the different white points.

[Doug Gray]

I would stay away from the Microsoft ICM engine. For instance it can show out of gamut colous converting from AdobeRGB to ProPhotoRGB because it does not perform chromatic adaptation related to the different white points.

Concur. Specifically, when converting using Abs. Col. ICM does not follow ICC's definition of Abs. Col.  but instead uses the CIE's definition.

Further, even with other RIs, ICM does a bad job converting and introduces significant errors in a few places. Adobe's ACE is far better.