Lately I’ve gotten quite a few “what’s the maximum ISO value I should use on camera X” type questions. I’m not quite sure what’s causing this, but I think it’s the wrong way to think about ISO.
The way to think about ISO is this: what’s the lowest ISO I can set without impacting my focus (aperture) and motion (shutter speed) decisions?
The problem with low light and bumping ISO is very simple: the randomness of photons is not your friend. Assuming a perfect sensor (no read noise, no thermal noise, etc.), the noise level is still the square root of the number of the photons converted. Capture 100 photons, you have noise of 10 and a signal to noise ratio of 10:1. Capture 10,000 photons, you have a signal to noise ratio of 100:1.
We don’t have a single camera on the market where raising ISO doesn’t have clear image quality implications. Well, that’s not quite correct. The Canon 1Dx, 5DII, and 5DIII, and the Nikon D3, D3s, and D5 all have lower ISO values that are constricted by read noise, so bumping ISO up above the base ISO to a modest level of 800 generally doesn’t have a problematic impact on visual quality on those models. So much so that I wrote in my book on the cameras that Auto ISO 800 is probably the preferred setting on a D3/D3s.
And yes, some older cameras—particularly early Canon DSLRs—tended to have razor-notched dynamic range charts where one specific ISO value was better than the neighboring ones.
Still, most cameras have a fairly linear relationship between ISO and image quality now, and this is likely to become more and more predictable as we move forward in the current sensor tech: ISO-less is the design mantra for CMOS at the moment.
What happens is that as you boost ISO you’re essentially becoming enslaved by that randomness of photons. The more you push ISO, the more you see noise caused by that randomness.
For almost two decades many of us shooters with technical knowledge of how the photo-to-pixel transition occurs have fallen towards defining what I call Usable Dynamic Range. Bill Claff calls his version of this Photographic Dynamic Range. His charts are easier to display here and prettier, so I’ll use his to illustrate the point.
Since many of these questions I’ve been getting center around recent cameras, such as the D500, I’ll use that chart here (use that link I just gave and choose your camera of choice from the menu if you want something different):
When you ask the question “what’s the maximum ISO value I can use on camera X” you’re really asking a different question that I can’t answer for you. That question is “what’s the minimum useful dynamic range I’d tolerate in my image?”
Bill uses a slightly different way of testing than I do, and DxO uses yet another evaluation method, but we each have one important element in our calculations: what’s the output?
Here’s the thing: no matter how good the rear LCD on your camera is, it will lie to you about “acceptability of images.” If you’re looking at the full image on even the best 3.2” display, you’ve buried noise so far by downsizing that you won’t see it. If you zoom in to 100% view, you’re looking at pixels in a way you normally wouldn’t see them in a print (or onscreen, even on a 4K display). So as noisy as it looks in the field at the pixel level, it may be just fine for your output.
And that gets me back to why I can’t answer the question: (1) I don’t know your output; and (2) I don’t know your tolerance for visible noise.
Personally, I try to gauge my own images based upon an implicit 13x19” print (maximum size of the desktop inkjets, and bigger than most client requests for my work are used at). It just so happens that my monitor, once I account for menu and scroll bars, etc., is very close to what would be a 13x19” print, so Fit in Width works fairly well as a stand-in for making an actual print). I’m less concerned about luminance noise than I am about chroma noise and loss of color fidelity (contrast changes). I’m willing to push my cameras to ISO values just shy of where I see those last two things clearly in a 13x19” print that I’m looking at from a slightly shorter than normal viewing distance.
But remember, low light is a very harsh mistress. I tend to like anything above a 60:1 signal to noise ratio. But in very low light I may never hit that, meaning that random photons are driving the propensity of noise and becoming visible. I’d love to make those little buggers non-random, but unfortunately my hair isn’t as white or unkempt as Einstein’s yet, and my brain lags further behind.
Ultimately, this brings us to exposure, because that’s what determines how many photons get to the sensor to capture. Exposure is the amount of light in a scene filtered by aperture filtered by shutter speed. ISO doesn’t come into play here. Light filtered by aperture filtered by shutter speed.
Thus, if noise is really the problem you’re trying to avoid, your best answers are not boosting the ISO. They are: (1) adding light to the scene; (2) opening up your aperture to a wider one; and (3) opening your shutter for a longer period of time. This is why many collegiate and pro indoor sports shooters rig strobes in the rafters of the arena (add light). That’s why pros like f/1.4 primes and f/2.8 zooms (don’t filter the light so much). That’s why it always pays to know what the longest shutter speed you can set that doesn’t show subject/platform motion is, and why IS/VR is valued when you don’t have subject motion (also don’t filter the light so much).
So here’s the 411 conclusion, in order of preference for most types of photography (some will flip 2 and 3):
- Boost light if possible.
- Open the aperture wider until you hit maximum.
- Lower the shutter speed until just before you have subject motion.
At the end of this process and only then you have to consider ISO boost (or underexposure, which does the same thing on many cameras, at least if you’re shooting raw, which you probably should be if you’re in tremendously low light).
ISO boosts are a last resort. And whether you’ll like the outcome or not is going to depend a lot upon how many photons got caught by the sensor (and with smaller or older sensors, often how much read and thermal noise is present).
