I've been struggling with how to explain ISOlessness in ways that most people can use to apply to their raw file photography.
I've taken another crack at it here:
blog.kasson.com/the-last-word/a-new-way-to-look-at-isolessness-2/
I'd be pleased to deal with comments and questions in this thread.
Great stuff. In a way we knew all that. But didn't connect the correlations and causalities to use this knowledge. Great! That's what turns people from denial to usage. I wish I could spend this time on audio HW issues. đ
Well, little problematic is that it needs knowing ones camera sensor. Not many people are into this depth.
Thank you .
That's a nice piece of work, Jim. I'm not sure that 'ISOlessness' is quite the concept it was when we all started talking about it, back when the D7000 came out. Since then cameras have become less 'ISOless'. The Z7 keeps on improving read noise all the way up (though I'm a bit suspicious about what that step down at 6400 is). A few comments - the legend on the Y axis didn't make much sense to me until I read it several times and worked out what you meant by it. Which is the problem, unless people are deep into this stuff, it's not going to mean a lot to them. I think there's a few general rules that might be helpful
- by and large, when RN goes below 2 e- you can probably forget about it for most purposes - so raising the ISO setting and losing a stop of DR just for 1/4 e- of read noise likely doesn't make a lot of sense.
- You won't get all the DR unless you expose right up to the FS - so in a sense ETTR does make some sort of sense above base ISO.
- DR is a moveable feast. Getting a stop extra DR by halving the read noise won't give you as visibly noiseless result as getting it by doubling the exposure.
- The real rule of ISOlessness is, if you're working in raw, there's a point where you might get a better result by fixing teh ISO setting and 'underexposing', subject to the characteristics of your camera, and what is the 'DR' or exposure range of the scene you're photographing.
Now it's more that cameras have ISOless ranges, rather than cameras being ISOless.
Agree. And that's what I think the graphs show. The point of the graphs is to show how clipping and noise change with ISO on the same chart, so people can see that the clipping losses as ISO increases are substantial compared to the noise floor reductions.
I'll work on that.
What I meant is that once you've fixed the exposure, cranking the ISO up to get the histogram on the right is not in general a useful strategy. There are some points in the ISO range for dual conversion gain cameras where it is useful.
This would probably make more sense to you: Log2 of photodiode charge. Then the noise legends could be modified to indicate rms noise. But I think that might confuse the technically unsophisticated.
That rang a bell - Sigma's first digital cameras were what could be called "ISO-invariant" - i.e., for a given exposure setting (=lighting, aperture, shutter), the image data written to the raw file did not vary with the ISO setting at all.
Then they introduced cameras with "AFE"s which placed programmable gain amplifiers between the sensor and the ADCs. That didn't last long. With the advent of the last "real" Foveon sensor, the 'Merrill', the ISO setting went back into the meta-data where it belonged đ
But then came the Fauxveon, the 'Quattro' and Sigma did a 180 and went back to AFEs.
A common whine with the Sigma AFEs was "blown highlights" with the punters not realizing the ADCs can get blown at any ISO.
Watch the wording. Cranking the ISO up to get the histogram on the right can still be useful. Because, arguably, you are not clipping the image if you do it properly, so there is no loss once you fit within the dynamic range of the scene.
I wonder though about handling it properly. More testing and discussion needed.
Yeah, it's hard to generalize without missing reasonable use cases. But the more I work on this, the more advantage i see to being able to properly deal with specularity in post production.
post comparison images and then everyone can figure out if its useful or not.
you need to account for what happens after "raw file photography" - in some raw converter... say developer ( ACR/LR is not the only out there ) tune some exotic NR parameters based on a nominal ISO recorded in a raw file, so all that good ISO is not part of exposure might notice-ably go out of the window...
That sounds like a terrible idea.
still it is possible, so you need to add always a remark - that while the theory is sound a practical application of a specific raw converter with a specific camera profile and other specific settings might render all that sound theory not useless of course but less useful... so user beware and actually DO TEST !
New labels in the legend, the y-axis, and the title. Better?
I should point out to others that 2e- is halfway between 0 and 2 on the y-axis of the plots I'm using here.
Why didn't you simply start the ordinate at 1e- instead of adding "above 1e-" to all the legends?
The y-axis does start at 1e-. Log2 of 1 is zero.
Blog post revised and corrected based on input from here and PS&T. Keep those suggestions coming. Thanks.
Most of the situations in which I wind up with very low exposure indices are ones where I can set the camera to M with auto-ISO and fixed "0 EC" (no ISO bias), and virtually never get any raw highlight clipping. So, I don't see any need at all to back off on the ISO setting to provide more highlight headroom.
The camera in question is very important, too. My Canon R5 has a visible fine horizontal banding noise that comes from post-gain read noise, and despite the fact that in an over-simplified mathematical sense the post-gain read noise should get lost in quadrature with pre-gain read noise, it clearly does not. Quadrature only applies directly to noise with the same spatial character, and standard deviation is inept to describe spatial character. I once synthesized two raw blackframes with exactly the same standard deviation and demosaiced them, and then downsampled them, and one turned black and the other one displayed a rainbow of large color blotches and bands.
So, in this case, the difference in the strength of the banding noise between ISO 25.6K with -1EC in 12-bit e-shutter mode and 102K with +1 EC inechanical shutter mode is quite large.
