THE CONSENSUS.
In early September, Andrew Rodney posted his own "real-world" example of
8-bit vs. 16-bit editing. As soon as it appeared, it was dismissed both by me
and by Lee Varis because it depended on an exotic RGB definition, the ultra-
wide gamut ProPhoto RGB, where the perceived impact of tiny variations is
much larger than in the RGB definitions used by almost everyone. Andrew
has known for at least five years that I consider testing in such RGBs
irrelevant--see "The Attempts to Obfuscate" below.

Even if we were allow Andrew to sneak this image in, it wouldn't show a 16-bit
superiority, for two reasons. First, he deliberately chose a zero-threshold
sharpening setting for his image to emphasize the differences between his
16-bit and 8-bit versions, when an equivalent setting that required no
additional steps was available. To back up the premise that you have found a
real-world difference, it's fair to say "I would like to do the following or
something equivalent, and I find that there is no way to do so without getting
an effect that displeases me." That would have been the case with any of the
four examples cited in Part II, "Where 16-Bit Can Be Better". It is, however,
fatuous to say, as Andrew does, "Of many possible equivalent settings, I have
intentionally chosen the one that displeases me, and now the test fails
because I am displeased."

Second, the standard is "totally obvious to anyone who looks," a "night and
day difference." Anybody who looked at the four examples cited above would,
in my opinion, prefer the 16-bit versions. The only person known to have been
shown printed versions of Andrew's samples preferred the 8-bit version.
Some areas of Andrew's 16-bit version are better than the 8-bit but larger
areas are worse. I think that a jury would rate the two versions as qualitatively
equal, but if they did not, I expect they would choose the 8-bit version as
better overall.

Andrew's edits were not particularly severe. Conceivably more massive edits
would have created a pair of images with differences strong enough for
people to have a clear preference for one as opposed to the other. In view of
my testing with B/W images, I think it's likely that when there is an actual
preference, people would tend to prefer the version corrected in 8-bit.
However, there would also be cases where the smoothing effect of 16-bit
editing would be preferable. But again, the situation is not real-world. Anyone
knowledgeable enough to take advantage of the extra bits would be unlikely
to be editing in such an RGB in the first place. And nobody, not even Andrew
and his partners, advocates making big edits in an ultra-wide RGB in
Photoshop proper. Any move big enough to provoke a visible difference
because of bit depth, they would urge making in Camera Raw.

Lee, Andrew, and I all agreed that when the file was put into Adobe RGB
rather than ProPhoto RGB, and the same edits, including the zero-threshold
sharpening, were applied, there was no significant difference between the 16-
bit and 8-bit version. Whereupon Andrew stated,

"I will admit: If you work with a small(er)gamut space, the need for high bit
editing is reduced and possibly a non issue."

I think that this face-saving admission is acceptable--that for Adobe RGB and
sRGB, the two spaces that we almost all use--editing in 16-bit is no longer
highly critical, no longer the distinction between a professional and a
recreational user, no longer a matter of night-and-day difference, but rather
"possibly a non-issue."

In return for Andrew's concession, I cheerfully respond as follows: "I
categorically recommend against making major edits to photographs in ultra-
wide RGB spaces such as Wide Gamut RGB and ProPhoto RGB when any
reasonable alternative exists. If you nevertheless insist upon doing so, you
should take into account that under such conditions the bit depth of the file
may affect quality, so you should be aware of the circumstances in which 8-bit
editing may be superior as well as those in which 16-bit may be."

THE ATTEMPTS TO OBFUSCATE.
Throughout these five years, the 16-bit advocates have been trying to deflect
attention from the basics, which are that they bullied and berated users,
calling them ignorant and unprofessional, all because these users did not use
an inconvenient workflow that has never been shown to be of any use
whatsoever.

The challenge is a simple one: to take any color photograph that might
conceivably be used in a professional setting, operate on it in any way that a
person, even an incompetent one, might try, and demonstrate a clear
superiority for 16-bit editing of that particular image. I do not say that no such
image could possibly exist. I do say that we have not seen one yet, and a lot
of people, including me, have tried hard to find one.

The terms of the challenge are extraordinarily permissive, and the 16-bit
advocates have resorted to extraordinary whitewashing to cover up the fact
that they can't meet them. They keep trying to sneak in obviously inapplicable
stuff; when everyone laughs at it they accuse me of "changing the rules" and
click their tongues and say, "see? No matter how hard we try, our stuff is
always rejected on a technicality! Clearly there's no image in the world that
Dan would ever accept!"

We saw this on the list last year. A member posted as an example a
Photoshop file consisting of no more than small blue area with noise. He
asserted that this was a picture of a "sky." There was no way of verifying this
as there were no clouds or other detail that might identify it as a sky; no way of
knowing that it wasn't computer-generated, for that matter. I said that it is
somewhat difficult to accept as a "real-world photograph" something that an
independent viewer can't identify as a photograph at all, let alone what it's a
photograph *of*. Andrew Rodney immediately went ballistic, again accusing
me of changing the rules and stating that this proved that I would never accept
any image as being real-world.

Similarly, the "real-world" corrections consisted of seven pairs of drastic
actions, each designed to reverse the other. Again, based on my assertion
that there was not the slightest chance that anyone would ever be stoned
enough to *dream* of doing something as crazy as that in the real world,
Andrew again asserted that there was a change in the rules.

In the 2005 ColorSync thread, I was again accused of changing the rules
because I should have made clear that "photograph" and "computer-
generated gradient" are two different items and that I would not accept a
gradient for testing. Also, it was alleged that I changed the rules because my
use of the words "color photograph" did not make it sufficiently clear that
grayscale images were excluded.

Andrew's recent image saw more of the same obfuscation. When I said that
the behavior of exotic RGB definitions is irrelevant, Andrew very piously said
what he usually does when caught cheating on such occasions: "I didn't
expect this scene or any additional scenes I can capture with this sub-$1000
camera to 'fit' your criteria since once again, that seems to be a moving target.
So I've come to the conclusion that no matter what one does to try to show the
advantage of high-bit editing, it will fall on at least two deaf ears."

I've never accepted such files for testing. Andrew Rodney knows this,
because he complained about this condition in 2000, in a thread that's
archived on our site. He correctly pointed out, and I agreed, that ultra-wide
RGBs are more likely to show a difference. I replied (remember, this is 2000):
"To me, it's like asking how many angels can dance on the head of a pin.
Nobody to speak of is using wide gamut RGB. There are a lot better uses of
our time than to investigate the question of whether it needs 16 bits. I *have*
looked at the issue in Adobe RGB and don't see that the extra bits are useful
there. As I mentioned earlier I would be anxious to see images from anyone
that might contradict this."

Further, in the 2005 ColorSync thread, I specifically suggested that the
ProPhoto space might sometimes show a difference between 16-bit and 8-bit
editing if the edits were extreme enough. Seizing that suggestion, Andrew
created his example--and then feigned shock that I wouldn't accept it, even
though he had known for five years that I wouldn't.

This type of behavior is only what we have come to expect. As has already
been pointed out, both Andrew and Bruce have tried to deny that they said the
wild things they did five years ago. And, particularly, the intentional
falsehoods on the Lindbloom site have been called to Andrew's attention
again and again on this list and elsewhere. Yet he continues to pepper
cyberspace with references to it. Every time he does so, he shames himself.

WHAT THE EXTRA BITS REALLY DO.
Understanding statistical concepts is not easy even for those with extensive
training in the field. Casinos have lost hundreds of millions of dollars to
mathematically astute players because the professors the casinos hired to
advise them on probabilities failed to comprehend certain statistical
interactions. The New York Times polled academics about a statistical-
interpretation puzzle called the Monty Hall Dilemma, and the majority of
professors of mathematics gave the wrong answer. Both of these examples
are absolutely child's play in comparison to trying to interpret what makes a
picture look good.

Progress in our field has been hampered considerably by "experts" who are
so terrorized by histograms that they don't understand that they feel they will
sound more authoritative if they try to terrorize their readers with them as well.

The calibrationism of the 1990s, the fear of applying easy curves to images for
fear of "irretrievable data loss", the phobia of LAB, the insistence on applying
profiles to things that obviously can't be profiled, are all symptoms of the same
disease. The confusion between gradients and digital photography is one of
the best examples yet.

A computer-generated gradient is a perfect file. Each pixel gets its optimal
value. Working in 16-bit gives extra precision, and it's unconditionally better
for gradients than 8-bit is. It's possible that the two methods might give
equivalent results. It's possible that the 16-bit might be better. But it's not
possible that it would be worse. Also, it's easy to come up with a
demonstration that shows damage by editing such files in 8-bit as opposed to
16-bit.

Now, contrast that to a digital capture. It is shot through a lens that is not
perfectly clean; through air that contains particulates; it is captured by sensors
whose performance varies with age and which may not be clean themselves.
The data is demosaiced by an algorithm that we know nothing about. The
vendors apply blurring to certain areas, and generally apply contrast-
enhancing curves in certain others. Finally, the data is written using an
unseen method to a format that may contain its own irregularities.

For all these reasons, the idea that even 128 levels per channel, let alone
256, are being captured with any degree of accuracy is very dubious. Yet if
you open a digital file, up comes what looks like a nice, smooth histogram.
That is enough for the 16-bit advocates to put their minds into neutral,
prostrate themselves in front of it, and start salaaming. They assume that
because the histogram looks like that of a gradient, the image will handle just
like a gradient does. That's like saying that because champagne tastes good,
motor oil must, or that because a helicopter can fly, it follows that a submarine
can.

In real life, parts of what the camera captures are much more reliable than
others. The better the shooting conditions, the more reliable the information.
An underexposed image has less than 8 bits of accurate information
anywhere. A shot in studio conditions has *more* than 8 bits of accurate
information in the midrange of its green channel, less throughout its blue
channel and in the shadows of each channel, and the rest is debatable.

CCD and CMOS devices, such as digital cameras, are notoriously poor at
extracting shadow detail. Every manufacturer tries to fight noise by using
some kind of blurring routine. Similarly, all digicams have difficulties with
noise in blue areas, such as skies, and result to all kinds of demosaicing
shenanigans to reduce it.

In these areas, the values that the camera spits out are approximations only.
In the darkest eighth of any channel, an 8-bit file has 32 available tones.
There is no digicam that I've ever seen that is even remotely close to
capturing these 32 tones reliably. An 8-bit file allows for more precision than
the camera has in shadow areas. In these shadows, the 8-bit file therefore
already consists of somewhat random numbers. Making it a 16-bit file adds *
totally* random numbers in these areas. The histogram-worshipers are fond
of using the term "quantization error", but this merely shows their statistical
naivete. In contrast to a gradient, where the 16-bit file is mathematically more
reliable, the 8-bit file is mathematically purer than the 16-bit in the shadows of
a digital photograph. Since, in shadows, the extra bits are random numbers,
working with a 16-bit file simply adds soft noise in transition areas. Skies
behave similarly.

My apologies for lapsing into techspeak. A reminder that this effect is so tiny
as to be undetectable under real-world circumstances. However, if the file is
stressed sufficiently (as Andrew was doing with his exotic RGB definition and
his sharpening sabotage) then we may be able to perceive it--and what we
will perceive is that editing in 16-bit amounts to applying a blur to these areas-
-a highly sophisticated blur that is hard to emulate by other means.

AFAIK, all of the claims about the images in which 16-bit supposedly is better
involve either skies or shadows. Makes sense--we often blur these areas
separately no matter how many bits are involved. The camera's algorithm
blurs them already, but sometimes it isn't enough.

Unfortunately, if you do enough to the 16-bit file that its inherent blurring has
an impact in shadows and in skies, it's likely that it also may have an impact in
areas where you *don't* want blurring. That accounts for the preference one
viewer had for the 8-bit version of Andrew's file. In the shadow areas it was
worse, but in the foreground areas that were full of detail, it was better.

An even more striking example showed up in the last two months, where
another 16-bit advocate posted a demonstration at http://www.visual-
vacations.com/Photography/16_vs_8.htm

The image is one of a barn constructed of dark wood, taken in normal
daylight. The complaint is identical to Andrew's, namely that correcting in 8-bit
created undue noise in the shadows. The procedure in this image is similar to
Andrew's as well, but there are some variations--especially in the result.

1) Both are definitely real-world images. This demonstration is in sRGB, which
is fine. Andrew's is in ProPhoto, which disqualifies it right off the bat.

2) This image is intentionally sabotaged by acquiring it in such a way that it
enters Photoshop grossly dark, and is therefore immediately disqualified,
because in the real world, we do not sabotage images on acquisition.
Andrew's image, by contrast, was acquired fairly.

3) Both images now take Levels changes and, in Andrew's case, Hue/
Saturation. Fine.

4) Now, the attempt to slip in the sabotage. Andrew sharpens using a zero
Threshold, which is clearly intended to bring out noise and thus disqualifies it
as a real-world move, since in the real world one could achieve the same
sharpening affect with a Threshold of 2 or even 1. The other demonstration
does this not just one better, but five better: it tortures the image with four
consecutive zero-threshold sharpens, plus two applications of Shadow/
Highlight, which is a form of sharpening. Again, not real-world: in the real
world, when we have two alternatives, we choose the one that looks better,
not worse. The purpose of sharpening is to make the picture look more
natural, not more noisy.

5) And the startling bottom line: while both exercises show more undesirable
noise in the shadows in the 8-bit versions, images don't consist entirely of
shadows. They need to be evaluated as a whole. I'm familiar with how juries
vote when given unidentified samples to choose from. We have one report
that with Andrew's image the 8-bit version was chosen as better. I don't think
that a jury would agree--I suspect the verdict would be a tie. With the other,
however, I have no doubt. The shadows are worse, for sure, but a jury would
pick the 8-bit version as better overall, because the same blurring effect in 16-
bit that helped the shadows also blurred the main interest object of the image,
the wood of the barn.

If editing in 16-bit were truly "better", there wouldn't be examples where
editing in 8-bit appears to give superior results--it couldn't, any more than
editing an 8-bit gradient can possibly produce a superior result to doing it in
16-bit. Also, it is very telling that the examples being used always are either
skies or shadows, and never the third major category of image that we
frequently have to blur: fleshtones. I would suggest that the reason is that
camera captures of fleshtones are very much more reliable than either of
skies or of shadows. There is therefore less need to blur due to inadequacies
of the original capture. Even though the extra bit depth in fleshtones is
probably real information and not noise, it doesn't serve a useful purpose.

OTHER VIEWS.
I would like to close by quoting at some length two other experts who have
conducted their own tests.In the 2005 ColorSync thread, Jeff Schewe
emerged after I had left, and, unbothered by the overwhelming evidence that
there is no quality difference at all, reiterated the same old silliness:
"I would argue that it [is] critical that ALL MAJOR tone and color moves be
done in 16 bit."

Jim Rich replied,
"As working in 16 bits to avoid all of these Urban Legend pitfalls such as:
The color being day and night different from 8 and 16 bit images.
That 8 bit files are fragile in terms of color reproduction.
Or that you will get artifacts like banding in gradients after a lot of image
editing.
That 10% percent of images require 16 bits so they don't break.
Or that you need to work in 16 bit because of device responses.
Let me say again on those points, I am more than skeptical that you require
16 bits. My experiences with RGB photos, CMYK print and Inkjet printing do
not indicate that one should jump to those conclusion. And we know that
mileage will vary.
And until I see a test or some other type of hard evidence that all of these
problems really exist and are as wide spread, as reported I can't buy into the
notion that there is an advantage to going to a full 16 bit workflow for a few
renegade images.
It would be interesting if some one would please show me a suite of images
where the benefits of 16 bit images jumps off the sheet and it is clear what is
an 8 bits and what is 16 bit file. Then I can really buy into the 16 bit way of life.
"

It's hard to know what more could be asked. Jim admits that it is impossible to
prove that there will *never* be a case where 16-bit is beneficial, any more
than one can prove that it would never be beneficial for a righthanded person
to work with the mouse on the left-hand as opposed to the right-hand side. As
indicated above, several 16-bit advocates have said explicitly that
demonstrations don't matter: it is so obvious that 16-bit is the better way that it
requires no proof.

Tim Grey, one of the rising stars in the Photoshop field, wrote this in his
September newsletter:
"In a general sense, I do indeed agree with Dan. The way I explain this
issue is that 16-bit editing provides more 'headroom' for editing, ensuring you
won't create posterization in the final result. However, I've also tried to make it
clear that the vast majority of the time, you will never see any benefit at all
from optimizing your images in 16-bit. The simple fact is, even if you strip out a
significant number of tonal values in your image, you'll still have an adequate
range to produce what appears, to our eyes, to be perfectly smooth
gradations. I often do a demonstration in my workshops as I posterize an
image and ask students to tell me when they see posterization. In general, I
get down to about 32 levels per channel before anything is visible, when
there are 256 available in 8-bit per channel images. So why do you need 16-
bit? The short answer is, if your computer resources can handle it, you might
as well keep your images in 16-bit to preserve as much detail as possible,
especially considering that output devices continue to improve, and at some
point we may have output that tests the limits of our 8-bit files, where you'd
actually see a difference.
The only time 16-bit would make a difference, for all practical purposes, is if
you need to make extreme adjustments to the image. If that's the case, you
probably would throw the image away, right? So 16-bit isn't a significant
advantage to most of us. Still something I recommend as a 'just in case', but
not likely to provide any practical benefit."

This is a sensible summary. Personally, I doubt that there's a realistic
possibility of benefit, but certainly you never know, and if you care to
speculate on it and have the disk space and computing time to spare, nobody
is criticizing. I believe it's more likely that some weirdness in an output device
will create a benefit than anything at the high-end. Naturally, I also think that
the most likely image to show an advantage would be one consisting of a sky
and little else, or an image with almost no detail outside of deep shadows.

LESSONS FOR THE FUTURE.
From the perspective of people who purport to be Photoshop authorities there
is a lesson to be learned about giving non-experts a hard time in public. You
never know when it's going to turn out that your views are mistaken, and then
you look ridiculous.

Second, admitting that one has been wrong is not a disgrace. Rather, I think it
gives one more credibility. Image processing is an infinitely complicated field.
In every new edition of Professional Photoshop I've pointed out areas where I
was previously incorrect. In the LAB book I talk about how I misunderstood
some of the sharpening issues. In at least two sections I pointed out images
that I had corrected badly.

Third, we have to grant that the difference in results between 8-bit and 16-bit
editing is not trivial to comprehend, even by people with a lot of color
knowledge, even after a lot of testing. OTOH, anybody who claims that
converting to LAB causes "catastrophic damage", that working on several
adjustment layers creates better data than applying the same corrections
consecutively, or that there is a "night and day difference, totally obvious to
anyone who looks" between 8-bit and 16-bit editing, is either too lazy or too
incompetent, or both, to do the ten minutes of testing that it would require to
cast serious doubts on these assertions.

Fourth, some sense of priorities has to be invoked. Some of the people who
have been most strongly in favor of 16-bit editing don't know how to set
highlights and shadows. Some actually edit not just in Levels but using the
master setting rather than channel-by-channel. Under these circumstances
the gain (if any) from 16-bit editing is about #4,807 on the list of things that
might help the image look better. Remember: even if some exceptional image
does show up that indicates an advantage for 16-bit editing, in all probability
the edge will be so minor that it can easily be compensated for. But to this
point, we haven't even gotten that far.

Users aren't blameless either. The entire imaging field is pervaded by
purported experts peddling solutions, myself included. Users tend to be
properly skeptical of claims made by unknown vendors, but a surprising
number of people are buffaloed by histograms and claims of mathematical
precision. Before buying into anybody's pet theories, readers should insist
upon images, not a bunch of pseudoscientific gobbledygook. If the speaker
can't phrase the concept in a manner you can understand, perhaps he can't
understand it either.

Users would be well advised to steer clear of anyone trying to justify
conversion or color correction theories with gradients, using terms like
"quantization error", or trying to convince us that a good-looking histogram is
more important than a good-looking image. Above all, when people come up
with new nostrums, ask to see images, not theories.

Finally, it should be remembered (as always) that color knowledge is always
evolving and that today's conventional wisdom is likely to be considered
wrong in ten years. I hope that the history of the correction method that started
out as "extremely critical", a "night and day difference," the difference between
a professional and a recreational user," and became "possibly a non-issue"
will be an instructive one.

Dan Margulis

[prev] [next]