A quick ISO invariance test - Alexander Becker

A quick ISO invariance test

Published: 08/11/2019

Apologies just right off the bat for click baiting with the milky way cover art, it seemed like it would be clearer than a plush toy :)

As photographers we often read about "ISO invariance" -- the idea that an image with the exposure boosted in post (i.e. lightened) will look the same as one just taken at the higher ISO (equivalent exposures). This is particularly relevant for landscape and astro, and to an extent wildlife photography too, especially with somewhat challenging birds to exposure for like a puffin for example.

I don't doubt this is a real thing, but have always been a bit skeptical -- in particular many of the tests I see online are using can labels or dollar bills. While these offer great contrast, they don't highlight noise that well and it's hard for me to translate noise in a dollar bill to something more "realistic" like feathers or fur -- the things I shoot on a day to day basis.

I finally decided to do a quick, and completely non-exhaustive, test. I photographed a plush kiwi bird toy to simulate a real bird, and then used a lamp to artificially light the scene. Our test subject is pictured below.

I choose ISO 1,600 as my "correct" exposure (shutter speed = 1/20, f/5.6) since this is the ISO level on my D500 where I start to think about noise and will apply noise reduction. In particular, I really pay attention to getting the exposure right once I get past this range, and tend to over expose the file a bit and pull it down a little bit in post.

Now that we have our highly cooperative subject and "correct expose" we can do our test. For each image I took two shots using mirror lock-up and picked the sharpest. This test isn't about sharpness, but given we are interested in how under (or over) exposure impacts detail it seems prudent. I picked ISO 100 (for 4 stops underexposed), ISO 800 (1 stop under), ISO 1,600 (goal exposure), and ISO 3,200 (1 stop overexposed). These images are shown below with no edits except for my default Nikon settings. Note that we're looking at 100% crops (one pixel from the image = one pixel on the screen) here as that is how to assess the true noise level.

In the ISO 100 and 800 images, the noise is minimal but of course underexposed and lacking in detail. The noise in the ISO 1,600 image isn't bad in my opinion, and the ISO 3,200 image has some noise showing, but could be handled relatively easily. 

However now I corrected each image to be the same exposure. This is where additional noise will be introduced for the for the underexposed file. In contrast, the overexposed file should (in theory) contain less noise overall as the signal to noise ratio will be optimized when corrected.

I initially thought that the ISO 100 file boosted 4 stops would be a mess, but it's not actually as bad as I thought. The noise in my opinion is worse and the details have suffered. In particular, look at the out of focus background and details around the eye. I can see a bit of a difference between the ISO 800 and 1,600 file, and same with the ISO 3,200 file. 

When the image is corrected with noise reduction these points become a bit clearer I think. I used Neat Image here since you can calibrate the noise directly. Since it's not a fixed value (like +1 or + 20), the noise reduction is applied with the same parameters for each image, but based on the individual noise level. The ISO 100 file is the worst (but it's not SO bad) and the other fiels are pretty similar to my eye with possibly a slight edge to the ISO 3,200 file.

Overall I was surprised by these results. When I underexpose files in the field I'm generally happy with corrected outcome, but I often kick myself for getting the exposure wrong. That won't change after doing this test, and since I do think the over exposed file does look a bit better, I'll likely keep doing that. But there are times you also want to underexpose to product particularly bright highlights and bring up the rest of the file in post, so it's great to see that the details in the file aren't too impacted either.

Of course, since this is a totally 'visual' test at just ISO 1,600 it's not exhaustive, but still good to see with your own eyes I think.

Feel free to follow me on Flickr or Instagram, or leave your thoughts below.