Last month, it appeared that some the basics were finally being agreed upon
as to some of the advantages or lack thereof in color-correcting 16-bit files vs.
8-bit files. This topic that has wasted far more time than it could possibly merit
over the last half-decade. Because there was a lot of posturing that masked
the essential agreement, I suggested that we stop the thread, and I promised
that at a later time I would post a fuller response.

I'd like to start from scratch and explain what the basic principles are, why the
debate went on so long, and what it teaches us for the future. I do not believe
this is a topic significant enough for a magazine article, but this post is article-
length, and is therefore divided into three parts plus an appendix. I intend this
to be my last word on the subject until the next edition of Professional
Photoshop, unless there are some new developments from images I am now
studying.

DEFINITION.
A bit is the smallest addressable part of a computer's memory. It can be seen
as either 0 or 1, either on or off, either yes or no. Each bit therefore has two
possible states. Two bits taken as a pair have four possible states. Three bits
have eight, four bits 16, and so on, with the number of possibilities doubling
every time a bit is added.

Ever since the advent of digital color correction in the early 1980s, it has been
standard to devote 8 bits of computer information to describe a single pixel in
a single channel. This gives a total of 256 possibilities. In three-channel
colorspaces like RGB, 24 bits (8x3) are required to fully define the pixel's
color, meaning that there can be 16,777,216 (256x256x256) possible colors
for a given pixel.

Nobody can see that many colors. Even the most optimistic estimates--and
that's the side you find me on--say that humans can only perceive a little more
than a million colors, and most experts put the figure considerably lower than
that.

Nevertheless, some believe that it makes sense to try to define more colors.
Some digital cameras try to record 10-bit (1,024 possible values per channel,
a billion possible colors). Modern drum scanners try for 12-bit (4,096 values,
69 billion possible colors). Whether these devices are accurate to that level of
precision is very doubtful.

In the mid-1990s, Photoshop introduced limited support for 16-bit files--
65,536 values per channel, 281 trillion possible colors. There was no
intermediate level. If you had a 10-bit file and wanted to get it into Photoshop,
you had to decide whether to bloat it by making it 16-bit or compress it by
making it 8-bit. Also, few important commands other than curves would work
on a 16-bit file. We could not even make a layered file in 16-bit. As time went
on more support was added, and today almost anything we can do in 8-bit we
can do in 16-bit.

THE ISSUE.
The question before us is not whether to capture in 16-bit, store in 16-bit, or
output in 16-bit. The only issue is, first, is there any advantage in *editing* files
in 16-bit rather than 8-bit, and if there is, it is so enormous as to constitute a
"night and day difference" or to justify the statement that anyone who does not
edit in 16-bit is a "recreational, rather than professional" user of Photoshop.

16-bit files are twice as large as 8-bit files. They take longer to store and to
back up and require more disk space. Also, if the file size is large, it may take
Photoshop much longer to perform edits. Plus, most output devices and many
layout programs won't accept 16-bit files, so we have to go to the trouble of
converting them to 8-bit eventually anyway.

For some, this isn't an issue. They process a limited number of images in
studio and, taking advantage of today's low prices, they have an infinite
amount of storage space. For others, like newspaper photographers on
deadline, doubling the file size would be so onerous as to be out of the
question even if there was a undeniable quality gain associated with it. For
everyone else, the doubled file size is an inconvenience to some degree. The
question must be whether we gain any benefit, and if so, what, because it's
possible we might want to use 16-bit some of the time and not others.

When someone advocates doing something inconvenient, whether
converting to LAB, or using Camera Raw, or doubling one's file size, or
putting 15 layers on a file, it's up to that person to make a compelling case for
it. It isn't up to you, me, or anybody else to show that it's *wrong* to do it. The
reason that the subject has refused to die is that the advocates of 16-bit
editing claim exemption from this rule: they take the position that whatever
they recommend, whether it's 16-bit editing or wearing garlic around the neck
while writing curves, must be taken as gospel and that they have no
responsibility to back up what they say.

AREAS WHERE EVERYONE AGREES.
Nobody AFAIK has ever doubted the following.

Extra bits are valuable in editing computer-generated graphics, especially
those that include gradients, or with image areas that are so heavily
retouched that they are essentially computer-generated as opposed to
photographic.

Scanners, digital cameras, and Camera Raw all operate natively with more
than eight bits. As there's no way to make them operate any other way, the
question of whether they *could* operate effectively with fewer bits is
irrelevant.

Certain scanners and certain camera software do not generate 8-bit files
correctly. Therefore, where possible, these files should be brought into
Photoshop in 16-bit, and converted to 8-bit at a later time.

Most people who handle lots of images have from time to time been burned
because they failed to save a copy of the original, untouched image. People
crop the image or rez it down only to discover that the extra information comes
in handy a year later. The chances of a bug in Photoshop or the OS
inadvertently damaging a file when it is resaved are extremely small, but they
are not zero. That alone is a good reason to save a copy of the original 16-bit
file.

16-bit does no real-world harm other than the extra space and computing time
it requires.

THE INITIAL WILD CLAIMS.
While nobody has ever argued that people who are comfortable using 16-bit
for correction should stop doing it, the same is emphatically not true of 8-bit
users. Starting in 1999, a slew of self-appointed experts, largely but not
exclusively Andrew Rodney and his business partners, began to attack
anyone who didn't use 16-bit for *all* editing. The rhetoric they used was
apocalyptic. Editing in 8-bit was "amateurish". It was "highly critical" to edit in
16-bit all the time. Those ignorant enough to edit in 8-bit proved themselves to
be "recreational, rather than professional" users of Photoshop.

Early on, a number of users questioned this, stating that they had compared
the two approaches found little difference in the results. Instead of accepting
the possibility that they might be mistaken, the 16-bit advocates dug in their
heels. I quoted the following from a single 2001 thread, where a squadron of
"experts" were berating their challengers. This is not a single speaker, but a
group of them, separated by ellipses.

"16 bit capability is critical during all aspects of tone compression…The
difference CAN be seen in the final output very easily. Most definitely on the
printed page, especially when using high-quality halftoning and even more so
to a film recorder…It's very easy to see that substantial color & tone editing
will eventually result in data loss and banding…If it means the difference
between taking a 16-bit image capture and editing that to the final image and
taking that same image in only 8-bit and editing that to the final image then
there is a difference like between the day and the night…Yes, if a histogram
full of holes has no impact on final output, then throw away the graphs and
just get on with the print run. However, all of us have Real World Output
showing the superiority of superior data acquisition…My advice? Take the
information you've read here to the bank. Stop doubting and start applying
what you've learned here…If you really start out with a RAW image in high-bit
form and a raw image downsampled to 8 bits, the difference really is night
and day. …it's totally obvious to anyone who looks that it's very advantageous
to do the big moves on high-bit data."

Nobody offered a single real-world image to show this enormous difference. It
was all histograms and gradients, gradients and histograms.

THE "CHALLENGE".
The type of color correction I teach does not depend on bit depth. The
methods work equally well whether you choose to use 8-bit or 16-bit. I have
never written an article or a column about bit depth. I don't even mention the
topic in my classes.

There are around ten pages of Professional Photoshop that discuss bit depth
and around five pages of Photoshop LAB Color. They are there is not
because they are necessary to my message in any way, but because the
advocates of 16-bit editing were so forceful in their denunciations of anyone
not using their methods that I kept getting hit with the question, both on this list
and elsewhere. At http://www.ledet.com/margulis/ACT_postings/
ColorCorrection/ColorCorrection.htm there are several archived threads on
this topic dating from 1999. The only thing that I could answer was that I have
nothing against others using 16-bit, and use it myself in dealing with
computer-generated graphics and in very specialized cases with color
photographs. However, AFAIK there are no real-world circumstances under
which a non-expert would find it beneficial for editing color photographs.

Before putting anything in my own books, I try to verify that there isn't
something unusual about my own files that causes me to draw an incorrect
conclusion. I therefore posted a request for people who thought that they
could demonstrate an editing superiority for 16-bit to arrange to send me files
for testing. As my own testing (see Part II, "Where 16-Bit Can Be Better") had
already established that a grayscale file could conceivably show an
advantage either for 16-bit or (more commonly) for 8-bit editing, I specified
color photographs only, in one of the four standard Photoshop RGB
definitions.

Around a dozen people have since responded between 2001 and now. They
put together packages containing proofs and often several different versions
of corrections. Over a period of a full week in 2002, I analyzed image after
image trying to find any circumstances under which 16-bit editing would give
superior results.

Finding none, in Professional Photoshop Fourth Edition, I published almost
ten pages of comparisons, because to illustrate the points, pictures have to be
fairly large. Six full pages were devoted to showing images at various
magnifications. Because the 16-bit advocates had retreated to a position of
claiming that the difference was only critical when the corrections were large
ones, the examples I showed ranged from large to inconceivably huge, in one
case taking a picture that was so flat that it was unrecognizable for subject
and correcting it into something that could be mistaken for professional work. I
printed each image at high magnification, including some individual
channels. They were printed without identification and readers were invited to
guess which was which. I particularly chose images that would be the most
prone to the sort of banding that the 16-bit advocates claimed would happen.

Every person who has ever submitted files to me has agreed with my
assessments of image quality. Certain people have had procedural mistakes
that I pointed out, and they have always agreed that my objections were
correct. In the cases where I am stating that there was a qualitative difference
when done in a certain way but not in another. I have shown the results to the
person submitting the files and they have always agreed with my findings.

Similar independent testing was subsequently performed by Jim Rich. The
main difference between his testing and mine was that Jim's testing involved
real-world corrections of images, in that his corrections, although severe,
might actually be seen on an everyday basis. Mine, OTOH, were intentionally
set up to be far more demanding than any real-world scenario would ever
entail. Anybody having to deal with the type of challenges that I published has
many more serious workflow problems than bit depth.

Jim's testing, which was also independently reviewed by experts, got the
same results. Since that time, around a dozen people have performed similar
tests, trying to find any real-world scenario in which 16-bit editing of color
photographs might produce an advantage, however trivial, over doing the
same thing in 8-bit. Everyone has come up empty.

In my LAB book, I had one further example, in trying to dispose of a similar
myth. In the mid-1990s, one of the same people who now fiercely defends 16-
bit editing was even more fiercely opposed to the use of LAB. He asserted
that the very conversion from RGB to LAB to do the editing caused
"catastrophic damage" to the image. This opinion was, of course, based on
analysis of histograms and gradients. In response, in a 1997 book, I showed
side-by-side images, one of which had been converted back and forth
between LAB and RGB 75 times. No difference, of course. Nevertheless, the
myth persisted, and I would several times a year get questions about the
supposed damage. So, in the LAB book, I did another such example, 25 times
back and forth. I compared it mathematically to other conversions and
demonstrated that the variation between the RGB>LAB>RGB and the original
RGB version was less than that between many RGB to RGB conversions.

Shortly after announcing the "catastrophic damage" theory and finding that
there was no damage at all even after multiple conversions, the theorists
changed their theory. The catastrophe, they opined, only occurred the *first*
time that a file was converted; subsequent conversions of the same file would
be harmless. But, if at some other point in the correction, there would be for
some reason another conversion to LAB, *that* would be a catastrophe. This
was similar to one of the changes in their 16-bit theory. Originally, it was
"highly critical" to do all edits in 16-bit. Then, it was changed to "big changes".
When it became clear that this theory didn't hold up either, it was changed to
"big changes done over a series of smaller changes."

Therefore, in total disgust, I spent five pages of the LAB book showing large-
size, magnified comparisons not of two variants, but of four, of each of two
different images. The images were specifically chosen because they had the
type of smooth areas that supposedly cause disaster in 8-bit editing and in
conversions to and from LAB. One was a 16-bit digicam capture, the other 16-
bit scanned film. They were compressed into a small 16-bit range, which is far
more challenging for the subsequent edit than starting with an original limited-
range capture.

To these images, I applied not one or two big corrections but, in accord, with
the theory, seven of them. These were done in RGB. I did the test once in 16-
bit, and once in 8-bit. But then I repeated the tests with a twist--after each of
the seven moves, I converted unnecessarily to LAB and back again.
Therefore, between the most politically correct of the four variants (16-bit all
the way, no conversions) and the least (8-bit all the way plus a conversion to
LAB after each move) there were seven night-and-day, totally-obvious-to-
anyone-who-looks corrections *plus* seven catastrophic-damage
conversions.

Nobody could tell which was which even at high magnifications.

Dan Margulis

[next]