Microsoft v9.

5
G Ga am mm ma a C Co or rr re ec ct ti io on n
Technical Memo 9
Alvy Ray Smith
September 1, 1995
Abstract
Gamma is one of the most misunderstood concepts in computer imaging ap-
plications. Part of the reason is that the term is used to mean several different
things, completely confusing the user. The purpose of this memo is to draw dis-
tinctions between the several uses and to propose a fixed meaning for the term:
Gamma is the term used to describe the nonlinearity of a display monitor. This
is the historical first meaning of the term. A corollary of this definition is that all
other uses of the term gamma should be dropped, or the imaging industry will
continue to stumble. In particular, this means that gamma should not be used to
describe nonlinearity introduced into the image data itself. Image data, especially
images intended for reuse, should always contain linear data, as commonly as-
sumed by all computer graphics algorithms.
Historical Gamma Definition
All computer display monitors (and all TV sets) are nonlinear. This means
that if the voltage on their electronics is doubled, their brightness does not double
as you might expect it to. In fact, it varies approximately as the square of the
voltage. If it varied by the square exactly, we would say it had a gamma of 2. In
other (mathematical) words, gamma is an exponent. But all monitors are slightly
different, so the actual gamma of a monitor might be anywhere from 1.4 to 2.6,
instead of 2. Very typical gamma numbers are 1.8 for the PC and Mac worlds and
2.2 for the broadcast TV world (and for PCs using TV graphics boards, such as
Truevision Targa+), but these should not be taken as gospel. They vary from dis-
play to display, even on displays from the same manufacturer. Any gamma other
than 1 is nonlinear.
The term gamma comes from the electronics industry. The equation derived
by electronics engineers to describe the intensity I, or brightness, out of a cath-
ode-ray tube (CRT) for a given voltage V applied to it its inputs is
I = kV

where the exponent of V is the Greek letter gamma (and k is some proportional-
ity constant)
1
. Most computer displays are still based on CRTs. So the historical

1
Equivalently, the intensity is related by the same exponent to the number of electrons in the
CRT that strike the phosphors on the face of the tube and cause them to emit light. This is because
the number N of electrons is directly proportional to the voltage V.
Gamma Correction
Microsoft Tech Memo 9 Alvy
2
meaning of gamma is that it is a number indicating the degree of nonlinearity of
a given CRT.
The proposal here is that thisgamma is a measure of the nonlinearity of a
display devicebe the only meaning that gamma have in imaging. Another use
of this term is discussed in the next section, then the difficulties resulting from
multiple, inconsistent uses of the term are discussed after that.
The important point to take from this section is that gamma is a property lo-
cal to each individual display device. Unfortunately, it is not a property that can
be obtained by software query of the display, at least with todays hardware. The
best that software can do, without input from the user, is guess a particular dis-
plays gamma.
Newer, Corrupted Gamma Definition
Many imaging applications today use the term gamma to mean a nonlinear-
ity introduced into the image data itself, independent of display of that data. This
is typically presented to the user by a curve that defines the nonlinearity to be
introduceda compensation curve, or tonal response curve, or correction
curve, etc. A common implementation lets the user fix the two endpoints of the
curve with one control per endpoint, then provides a third, central control called
gamma that controls the curvature of the central part of the curve. The result of
applying this curve to an image is that the original pixels in it are replaced with
new pixels containing the nonlinearity described by the curve. The details of the
implementation need not concern us here. The important point is that the term
gamma in this context is used to mean a nonlinearity introduced into the pixel
data itself. Notice that this means the nonlinearity is global to all display devices
(on which that image might be displayed). Equivalently, the nonlinearity of a
display device is compounded by the nonlinearity introduced into the data itself
when it is presented to the user.
It is clear why this corrupted definition arose. In both cases, gamma is a sin-
gle number or exponent defining the curvature (nonlinearity) of a curve. In both
cases, the apparent brightness of the display image is altered by changing this
exponent. So what is the problem?
The Assumption of Linearity
The problem is that all computer graphics computations assume linear im-
ages. This means simply that, for instance, half red plus half red gives full red.
This is fundamental to the industry. For example, the composition of images uses
linear interpolation of images as the algorithm, which assumes the images being
composited together are linear. Another example is antialiasing of edges to rid
them of the jaggies. The classic computer graphics solution assumes linear data.
All nonlinearity must be removed before entering the computer graphics world
and then reintroduced, if necessary, when exiting it.
There are two ways to take care of this mismatch between the reality of non-
linear displays and the assumption of linear computations:
Gamma Correction
Microsoft Tech Memo 9 Alvy
3
1. Take care of the nonlinearity in the display, or
2. Take care of it in the data.
Only choice 1 preserves image data for later usefor reuse. This is why gamma
in the display context is proposed as the only use of the term. As you can proba-
bly foresee, I am going to argue an even stronger point: Nonlinearity should
never be stored in an image. Or, if it is, then this nonlinearity must be noted in
the storage format in such a way that it is known how to remove it to retrieve
linear data.
Consequences
Unfortunately, many applications force you to take care of display nonlinear-
ity by making image data nonlinear. They do this by assuming the default moni-
tor gamma is 1, or linear (which, of course, it is not). The user then introduces the
nonlinearity necessary to make the image look right on the monitor by varying
the applications brightness compensation control, often unfortunately called
gamma. This replaces the images of the pixelwhich hopefully are linear by
defaultwith nonlinear pixels. This works, so long as you never use the result-
ing image for another image computation, and so long as the next display you
show the image on (including ink on paper) has the same nonlinearity as your
original display. This has "worked" often enough or closely enough in the past
for the mistake to have been tolerated, but it is only an accident when it works.
To state the problem another way: The mistake works so long as monitors of
identical gamma are used for any single image and so long as nothing further is
done with the image, computationallyie, it is only displayed thereafter on the
monitors of the given gamma. You can see why the problem is rampant: This set
of conditions is often met in the case of a single user generating a single image
for no future reuse (computationally).
But the whole idea of an image composition program (Altamira Composer,
for example) is to take images from many sources and composite them together
to form new images (which may then be used by someone else as a component in
yet another image, and so on). Images or sprites are reused; reuse is a fundamen-
tal of modern multimedia content creation. Such an application must assume lin-
ear data in its images for these computations, and hence nonlinearity must be
handled by the display.
Consider this: Kodak PhotoCD images have a gamma of 2.2 embedded in the
image data
2
. This assumes that all monitors they are to be displayed on have a
gamma of 2.2. But we know that many popular monitors for PCs have gammas
of 1.8. And for applications, like Altamira Composer, that have to assume linear
image data, the PhotoCD images are nonlinear and must be de-compensated be-
fore they can be used in the application. Unfortunately, Kodak does not supply a

2
To be careful with my terminology, I should say here that Kodak has embedded a nonlinearity
in PhotoCD data equivalent to displaying linear data on a monitor with gamma 2.2.
Gamma Correction
Microsoft Tech Memo 9 Alvy
4
programmers interface (API) that allows the linear data to be extracted from its
format. So computer graphics applications have to sacrifice 1-2 bits per channel
per pixel to get linear data. Furthermore, an applications programmer must
know that Kodak has assumed 2.2. This fact is not stored in the image file. I
learned it by reading the documentation that showed the encoding formulas
used by Kodak. This documentation shows exponents of value .45 in several
places. I happen to know that 1/ .45 is about 2.2 so could easily recognize the em-
bedded gamma correction for what it was. I advocate, of course, that Kodak not
correct (delinearize) its image data, but since it is too late presumably for them to
change their format, then they should provide an API that allows developers to
decorrect (linearize) their data with as much accuracy as possible.
Consider this: A user exports an image from Photoshop on a Mac (gamma
1.8 nominally) to Photoshop on a PC with a video display (gamma 2.2 nomi-
nally). In general the PC image should look too bright, so the user will either de-
brighten the display, putting it into who knows what state, or will recompensate
it in the imaging app, thus introducing a new nonlinearity into the data and pos-
sibly throwing away some bits. This happens very often. This assumes Photo-
shop is used in its default way: Monitor gamma is assumed to be 1 and nonlin-
earity is wired into the pixel data
3
. I am not just picking on Photoshop. Microsoft
Imager does the same thing. It has, in fact, a slider labeled on its main display
that affects image data, not the display.
The Altamira Composer solution to the gamma problem assumes the default
monitor gamma is 1.8 for PCs. This can be changed to a particular displays
gamma in the View Options menu. The application corrects for the nonlinearity
of the display during the display process, not the computation process. The data
in images inside Altamira Composer can always be assumed to be linear, and
linear data is always stored to files.
If images are imported to Altamira Composer from applications that have
forced the display nonlinearity into the data, a user willnot surprisinglyget
unexpected results. For this reason, Altamira Composer provides for the removal
of the unwanted nonlinearity with a so-called "degamma" process. Of course,
there is no way for Altamira Composer to know exactly what nonlinearity has
been stored in the imported data. (That is the problem with storing the nonlin-
earity in the data.) It just guesses by using the monitor gamma setting in View
Options to degamma the data. So for nonlinear images imported from a broad-
cast video app, a user might want to set this option to 2.2.
Another common source of nonlinear image data are applications supporting
the Targa (.tga) image format and Windows (.bmp, .dib) image format
4
. Altamira
Composer provides an option called TGA, BMP Degamma Correct in File Op-

3
There are ways around this in most imaging applications, but unfortunately they are not the
default ways so ordinary users never discover them.
4
The Windows image (bitmap) format is apparently derived from the Targa format, so it is not
surprising that they share this problem.
Gamma Correction
Microsoft Tech Memo 9 Alvy
5
tions to automatically take out the nonlinearity in these images. It defaults to No
(since Altamira Composer itself stores linear data in these two formats). If images
are imported from applications storing nonlinear data in these formats, this op-
tion is set to Yes. Caution: Targa apps are usually for broadcast TV with typical
gamma of 2.2 (or even higher). Windows apps usually assume 1.8.
Gamma can be confusing, as the above probably illustrates. Here are the
simple rules Altamira Composer uses and what I am advocating that imaging
applications do as a matter of course: Images are always assumed to be linear.
Gamma is applied only to the display of images and not to the data of the im-
ages. The display
5
is assumed to be nonlinear (because it is). Applications sepa-
rate computation from display cleanly, and gamma correct for the local display
only in the display process.
To get compatible results between imaging applications written under the (I
trust you believe sensible) new guidelines offered here and those written the
old way: Set the monitor gamma assumption in all the old imaging applica-
tions to the same (greater than 1) valuepresumably to that matching ones
usual display monitor. Most applications provide a way to do this. This transfers
the nonlinearity correction in those apps from the computation process to the
display process, as it should be, leaving linear data in the images themselves.
A desirable consequence of all this is that it would be very convenient for
imaging software if display devices provided gamma correction tables settable
by software. That way, each imaging app could work completely in linear space,
knowing that the display step would be correctly compensated by the local
monitor for its local nonlinearities. Believe it or not, this was the way it was done
20 years ago, but the idea got lost along the way, leading to the mess described in
this memo. Unfortunately, it is probably too late to change. The technique of-
fered here is the best that can be done short of changing all the hardware.


5
As is probably clear, I include as display ink on paper, video monitors, flat-panel displays,
and photographs. They are all nonlinear in different ways.