Note: this post has been revised and corrected based on input from DPRevived and the PS&T forum on DPR. Thanks to all involved.
Recently, I wrote two articles for Lensrentals on raw exposure.
One of the takeaways from the articles was, with modern cameras, by far the biggest win associated with exposure to the right (ETTR) occurs at base ISO. At the same exposure, there is little return in increasing the in-camera ISO setting to push the histogram up against the righthand side. I presented a handwaving argument for that, and a technical one, too. But I sensed that I’m not getting through to everybody. So, I’ve been looking for a better presentation.
Before I dive in, I should mention that Bill Claff has done some good work in this area. Once you learn how to read them, Bill’s curves are very useful. But understanding what’s behind the curves is a bridge too far for some people. First, you have to get your head around how photographic dynamic range (PDR) is calculated. Then, you need to understand the machinations that lead to the shadow improvement curves. If you’re technically sophisticated, it’s not hard. If you have a hard time remembering your high school algebra, it’s tougher.
Let me present what I’ve come up with, and I’ll walk you through it.
The above set of curves is applicable to a Nikon Z7. Data was obtained by dark-field images with electronic first-curtain shutter (EFCS) at 1/1000th second. The vertical axis is log base 2 of the electron count. If you’re adverse to logarithms, just think of it as sensor response in stops.
This presentation reflects a basic fact about CMOS sensors. There are photodiodes associated with each pixel. The photodiodes react to photons falling on them by producing free electrons. The photodiode response to light, measured in electrons at the photodiode for a given light intensity at the sensor, is unaffected by camera ISO setting, except that higher ISO setting can induce clipping. The graph shows how changing ISO setting affects the upper limit of how much light can be recorded, and how the effective lower noise limit changes as well.
In the above curves, there is no normalization for sensor resolution; Bill’s PDR curves are normalized. The red, green, and blue lines are the read noise referred to the sensor photodiode itself. They can be thought of as the noise floor. You will be able to make out details in the image below that floor, but they will be extremely noisy and, in my opinion, not useful for anything approaching general photography. The opposite end of the dynamic range is captured by the black line. Light levels above that will not be captured at all, but clipped to that value. The engineering dynamic range (EDR) at each ISO setting is the vertical distance between the black line and the red, green, and blue lines.
Take special note of the way the read noise declines at ISO settings above 5000. This is not a welcome occurrence, since it is the result of Nikon’s digital processing of the raw files. I would suggest avoiding those settings unless you have no alternative. However, the reduction in read noise upon transitioning from ISO 320 to ISO 400 is real, and is the result of the dual conversion gain architecture of the Z7 sensor. You can see that the read noise is essentially flat from ISO 64 through ISO 320, and from ISO 400 through ISO 5000. The common term for those regions is “ISOless”. You can see what happens to the EDR as you increase ISO setting in those flat regions: read noise stays about the same, but full scale drops, leaving you with less EDR.
But there’s another kind of noise that affects the shadows in images. It’s called photon noise, or shot noise, and it is an inescapable result of counting photons, which is what the sensor in your camera does. I’ve added curves in magenta that indicate the sum of the photon noise and the read noise for three different signal levels.
The highest signal level is graphed with a dashed line. That signal level can be plotted on the graph as a horizontal line at the vertical axis tick labeled 8. That’s about 8 stops below full scale for ISO 64, and with a conventional exposure where a 100% matte reflective surface is full scale, would correspond to about a 0.4% matte reflective surface, or about 5 stops below middle gray. You can see that the read and photon noise at that signal level is virtually a straight line: ISO setting makes no difference. I probably should have stopped the high signal-level noise curve at ISO 8000, since after that it is above full scale.
The next highest signal level is plotted with alternating dash lengths. That signal level can be plotted as a horizontal line at the vertical axis tick labeled 6. That’s about 10 stops below full scale for ISO 64, which is very dark. You can see some changes in the photon plus read noise line there, but they are trivial compared to the loss in the highlights as you increase the ISO setting.
The lowest signal level is plotted with short dashes. That signal level can be plotted as a horizontal line at the vertical axis tick labeled 4. That’s about 12 stops below full scale for ISO 64, which would correspond to a stygian matte reflective surface. At that level, there is a useful, but small, improvement in dynamic range when increasing the ISO setting from 320 to 400.
Here are similar curves for two more modern cameras.
Except for the transition from ISO 64 to ISO 100, the Hasselblad ISO steps are full stops, which is enough that the change from ISO 100 to ISO 200 offers no improvement in dynamic range even for the lowest signal level plotted in magenta.
What’s the take-home lesson? Increasing ISO setting in-camera decreases headroom much faster than it lowers shadow noise. Be careful when cranking up the ISO.
Caveats (warning, some of these are technical):
This post only applies to raw shooters. The rules are different for in-camera JPEGs.
Some cameras, at some ISO settings, don’t apply gain to the raw signals before writing the data to the raw file. They expect the raw developer to do that. The Fujifilm GFX 50S and GFX 50 R are such cameras, but the GFX 50SII is not such a camera.
Modern CMOS cameras use a combination of methods to apply gain to the charge on the photodiode. One method is to change the size of the capacitor associated with the sensor pixel, so that the same number of electrons results in different voltages at the input to the source follower. Another is to change the gain applied to the signals from the source follower before the analog to digital converter (ADC). This is usually done with a device called a programmable gain amplifier (PGA). A third way to increasing the raw values written to the file is to apply digital multiplication to the output of the ADC. Some cameras use all three of these methods at some ISO settings. The ADC is designed to clip before the voltage at the capacitor associated with the photodiode becomes significantly nonlinear. So all three methods result in the same clipping at full scale shown on the plots above.