The high ISO capabilities of cameras seem to fascinate many photographers. Unfortunately, there are many misconceptions about this subject. I undertake an attempt to make the subject understandable with detailed demonstrations.
The demonstrations are based on images created by a Nikon D3, but everything applies to many other cameras. The analysis was done on unadultered raw data, by
Rawnalyze.
The highest "true" ISO setting for a few other cameras:
- Nikon D300: 1600
- Nikon D40X: 1600
- Canon 40D: 1600
- Canon G9: 800
- Sony A700: 2500 (?)
- Olympus E3: 3200
Six images are used in the demo: one ISO 6400, one 12800 and one 25600, all in 12bit and 14bit. The setting was the same, the results are comparable.
There was an important difference between the two sets, more important than the bit depth. The 12-bit set was created with different exposures (1 EV lower with ISO 12800 than ISO 6400, and two EV lower for ISO 25600); thus, the histograms of these shots look identical, at least at first sight.
The 14-bit set was created with fixed exposure, measured with ISO 6400, thus the 12800 and 25600 images are much brighter than the ISO 6400 image.
(The images used in the demonstration are courtessy Albert C. Lee, a fellow poster on the DPReview forum and the owner of a D3, unlike me.)
For those interested: before reading on, pls download following TIFF, as I will refer to the 25 or so layers in it. File size: 6 MB.
Screen captures in layered TIFF The speciality of "very high" ISOs
The main difference between ISOs is, of course the gain: how the electric change is converted in a numerical value; another difference is, how reliable the result is, i.e. how "clean" the resulting data is (noise).
And then there is a little known difference: if the gain with an ISO is really different from the gain of other ISOs. The answer is in short: not always.
Specifically relating the D3: the highest ISO, which yields a different gain from the lower settings is ISO 6400. The higher settings are in reality equal to ISO 6400, but the resulting pixel values are numerically manipulated. Actually, the same is true relating to the 1/3 EV settings as well, but I have not analyzed these with the D3.
Before anyone gets upset either with this fact, or with me stating it: there is nothing new on this, most cameras are acting so. My Canon 40D goes up to ISO 3200, but ISO 1600 is the last true ISO, and all 1/3 EV ISOs too are numerical derivatives.
I will refer to those ISO settings as "fake ISO".
I am pretty sure some (or many) of the readers would immediately ask for proof, so I rather demonstrate the situation now.
The layers are in three groups; now I am referring to the group Histograms. There are two screen captures for each shot, both show the *raw* histogram of the shot. The "coarse" histogram shows the entire range, 0-4095 or 0-16383. The histogram consists of 512 columns; accordingly, each column represents an 8 or 32 wide range of pixel values. However, there is one column for each pixel value in the "fine" histogram; the range of currently displayed pixel values is shown as well.
We are dealing now exclusively with the fine histograms.
Let's see first ISO 6400, 14 bit. Pls magnify the image in PS, so that you can mark an area and count the number of columns.
There are "gaps" (black columns) in the histogram; their meaning is, that there are no pixels with the respective pixel values. When this occurs in the middle of a large "bulk", like here, then this fact has a special meaning: that pixel value is "spare".
Take a look at the magnified red histogram. Mark a repetitive range of gaps and count them. I see 17 gaps in a 100 wide range, and this remains so in higher ranges as well, with some small disturbances. Thus, we can say that the proportion of effective pixel levels is about 83%. This translates to about 13600 effective levels from the 16384 (don't worry, they are more than enough). Actually, these numbers changes slightly with the firmware version, but this scheme applies to all lower ISOs as well.
The blue channel is a bit different, the proportion of the effective pixels is about 92%.
Forget about the green channel, it seems to be "stirred up".
Now, let's go to the ISO 12800 image. The proportion of effective red levels is about 43% and of the blue ones is about 46%.
The ISO 25600 image yields about 21% for the effective red pixels and 23% for the blues.
What does thie above mean? It's simple: the pixel values gained from ISO 6400 get multiplied by two and that becomes ISO 12800, multiplied by four it "yields" ISO 25600.
This trick can not be observed on the 12-bit images so clear, because the number of levels are enough to populate the range better, even after the multiplication by two. However, the top edges of the histograms are telling, and the histogram of ISO 25600 shows the gaps clearly.
This is the reason I call these "fake ISO".
High ISO and noise
It is a very common misconception, that high ISO causes noise. The truth is, that not the ISO is the cause of the noise but the low exposure; although this often goes hand in hand with high ISO, this differentiation is very (but very) important.
Let's see now the noise in the demo-images.
Notes:
1. the color of the crops is off, for
a. Rawnalyze, the program I am using for the analysis does not carry out demosaicing and color conversion from the camera's color space in sRGB,
b. no WB has been carried out yet, for I wanted to keep the raw channel related values intact even in the RGB statistics,
Furthermore, the colors from the 12-bit and 14-bit images are way different due to different light source,
2. the relevant difference between the 12-bit and 14-bit set is *the exposure*; the bit depth does not play any role here, except in the appearance of the fine histogram.
So, let's compare the noise.
First, compare the exposures on the coarse historams of the 12-bit shots: they are practically identical (of course, as the shutter was halved with doubling the ISO). Based solemnly on the histogram, the images ought to look quite the same.
Disable the Histograms layer group and open the Noise group, comparing the layers "ISO 6400 12bit noise" with "ISO 12800 12bit noise" and "ISO 25600 12bit noise". The higher noisiness of the higher ISO images is not only clearly visible, but it can be quantified on the RGB statistics: the third number in each of the three number groups is the average RGB pixel valu; these remain relative consistent between the images. However, the fourth number, the standard deviation increases sharply with the ISO.
(These statistics values relate to the small selection, marked by an orangy recangle.)
The reason for the noise can be seen in the "ISO 6400 12bit to16" etc. layers: the strangely colored pixels are those with values under 16.
I chose 16 as pixel value limit for "noise" with 12bit depth, and 64 with 14bits. This choice is not completely arbitrary: these are the "borders" to the 9th EV from the maximum exposure downwards.
Anyway, the number of pixels in the "noise zone" increased drastically with the higher ISOs. I.e. the noise is the product of low pixel values with high variations (here quantified by the standard deviation).
Now, check out the coarse histograms of the 14bit images. (I feel compelled to repeat: this is not a 12bit vs 14bit issue!) The histograms reflects the fact, that the aperture and shutter were the same in all three shots. And the noise? Virtually equal in all three shots, with close the same proportion of "low" pixels - despite the different ISOs.
So, do higher ISOs cause the noise or the low exposure?
Dynamic range
I hope to have made it understandable by now, that ISO 12800 and 25600 on the D3 do not contribute to the *raw* image data at all. However, is that all to know?
Using the fake ISOs is not only not advantageous, it is doublessly disadvantageous, because the "numerically improved" pixel values take up space in the numerical range. This sounds a bit abstract, what does it mean?
First the demonstration: pls turn on the Clipping layer group and off the others. There are three captures there, showing the same segment of the image.
One capture is with ISO 6400, 14bit. The clipping indication is turned on (the checkbox "Raw clipping"), but there is no clipping here, apart from a few stray pixels (and those are clipped at the dark end). The DoF was not particularly large, but some details are visible in the middle of the crop as well.
Now, look at the layer "ISO 25600 14bit highlights gone": some of those details are not perceivable any more. Why? Look at the next layer, showing the clipping. Magenta indicates the lack of green, and the red shows, that both the green and the bluwe pixels clipped there.
So, using ISO 25600 not only did not contribute to the image quality, but it damaged that: it caused clipping of some areas. 12800 decreases the dynamic range by one stop and 25600 by two stops, compared to 6400.
Conclusion
Using the fake ISOs should be limited to cases, when the exposure can not be increased by any means, AND either JPEG image is created in-camera, or the brightness of the image on the LCD is crucial (why ever).