We’ve seen how information about a photographic scene is collected in the ISOless/invariant range of a digital camera sensor, amplified, converted to digital data and stored in a raw file. For a given Exposure the best information quality (IQ) about the scene is available right at the photosites, only possibly degrading from there – but a properly designed** fully ISO invariant imaging system is able to store it in its entirety in the raw data. It is able to do so because the information carrying capacity (photographers would call it the dynamic range) of each subsequent stage is equal to or larger than the previous one. Cameras that are considered to be (almost) ISOless from base ISO include the Nikon D7000, D7200 and the Pentax K5. All digital cameras become ISO invariant above a certain ISO, the exact value determined by design compromises.
In this article we’ll look at a class of imagers that are not able to store the whole information available at the photosites in one go in the raw file for a substantial portion of their working ISOs. The photographer can in such a case choose out of the full information available at the photosites what smaller subset of it to store in the raw data by the selection of different in-camera ISOs. Such cameras are sometimes improperly referred to as ISOful. Most Canon DSLRs fall into this category today. As do kings of darkness such as the Sony a7S or Nikon D5.
Refer to the earlier article for a more in-depth description of the diagram above and remember that this is a simplified model chosen for ease of getting the point across. Many cameras do not actually work like this, but in many cases this model represents a decent and intuitive approximation.
Non ISOless: Downstream DR Lower than Sensor’s
For a given exposure this is the path that faces scene Information collected in the non ISO invariant range of a digital camera on its way to being converted to digital numbers and written to the raw file.
Contrary to the situation in the earlier ISOless case, where the noise floor* of each stage was the same or lower than at the photosites, by the time the signal of this imaging system as set up reaches the ADC and related circuitry the noise floor is two stops higher than at the photosites. This means that for a given Exposure the photosites can collect more information about the scene than the electronics are able to process into raw data in one go. This effectively reduces the information carrying capacity of this system by two stops from 14 to 12 bits (photographers would call it Dynamic Range, in stops).
Same IQ as ISOless in Highlights
At base ISO such a system records highlights like its ISOless counterpart does, all else being equal. But as the Signal dips into the shadows the camera’s noisier ADC input results in SNR worsening faster. By the time the Signal gets into the deepest shadows random digits are encoded into the raw data where information from those last two stops should have gone. The system is not able to transfer the last two stops of tonal information from sensor to digital numbers properly, so while ‘Dark Detail’ tones in the diagram above are acceptably present at the photosites they are gone forever by the time the relative information is written into the raw data.
Tone Selection by ISO
Since deepest shadow tonal information is present at the sensor’s photosites, however, the photographer can choose to amplify it before it gets to the ADC’s noisy surroundings so that it is no longer below the noise floor. They do this by raising ISO, in this example by two stops. As explained in the earlier article, if ISO is raised with a fixed Exposure highlights are lost to clipping stop for stop, so the DR of the resulting information will still only be 12 stops. But at least the photographer will have chosen what 12 stops out of the 14 available at the photosites to save in the raw data. This is what raising ISO from base to 400 does to tone information transfer:
The entire Signal out of the photosites is increased by two stops, including related noise. Therefore the Signal to Noise ratio (SNR) out of the amplifier remains the same or similar as at the photosites. In so doing, however, the top two stops of highlights were pushed beyond the ADC’s range, clipping them off. But at the same time the desired ‘Dark Detail’ information was encoded with an acceptable SNR, therefore being saved usefully into the raw data. In addition every other tone from the photosites was amplified, therefore increasing the Signal relative to noise at the input of the ADC thereby improving relative SNR in the shadows.
Improved SNR in the shadows: As at the Sensor
That’s why when, for a given exposure, one raises ISO in a camera in a non ISO-invariant range the deep shadows recorded in the raw data appear cleaner: they have higher SNR, better IQ. In fact what the photographer has done is made a conscious choice to capture in the raw data the better SNR available at the photosites in the deep shadows, giving up in exchange highlight headroom.
We could have raised the ISO to 200 only, thus compromising just 1 stop of highlights but then saving just one stop of shadows, final DR unchanged at 12 stops. And you can see that in this example raising the ISO above 400, say 800, would have lost three stops of highlights but only netted two stops of SNR improvement in the shadows, thereby reducing DR to 11 stops. In other words, this exemplified camera is ISOless above ISO 400, where raising ISO reduces highlights stop for stop without providing an improvement in SNR.
Information about the scene (IQ) is best right at the photosites and can only be degraded from there by imperfect information transfer. Ideally all properly designed** imaging systems would be ISO invariant, that is capable of storing all scene information collected by the photosites during Exposure into the raw data. That typically requires short paths from pixels to ADC, which today means bringing the ADC on board the sensor chip forcing compromises elsewhere (for instance readout speed). We can’t have our cake and eat it too, yet.
* I know this is overly simplified, but follow along for the sake of clarity.
** In a well designed imaging system the encoded bit depth needs to be maintained higher than the ratio between the capacity of the ADC and read noise at its input, in the same units, expressed as a power of two logarithm.
Only the manufacturer can measure the actual noise level at the input of the ADC. What we can estimate instead thanks to Photon Transfer Curves is the random read noise referred to the output of the photosites in physical units of photoelectrons (e-). If analog amplification and transfer of the e- to the ADC adds little noise, we can assume that the estimated noise out of the photosites is about the same as that at the input of the ADC. That is not always the case. This subtle difference can sometimes result in interesting PTC responses near base ISO for overly clean sensors, even with an estimated input-referred read noise larger than 1 DN/LSB (latest Exmors, see for example some curves at Jim Kasson’s).
What a wonderful blog this is. So much information presented in a very lucid manner. Thank you!
My pleasure, Mike, thank you for dropping by.
Jack
Hi Jack !
I’m a french photographer and I just found website, searching about iso invariant/iso variant, etc…
What a great website !!
I’m far to unterstand everything but I have a lot of curiosity about all of this !
I have a question, maybe it’s super easy for you but I’m not an expert (not yet :D) : how can I know what type of sensor have my camera (iso invariant/iso variant/not totally invariant…) ?
I mean without someone telling me what it is, is there a way to find by myself and learn ?
I have a Fuji X-T1 🙂
Thanks again for all your precious informations, I’ll read all of your website and try to learn as much as possible 🙂
Best Regards,
Frankie
Hi Frankie, thanks for your kind words. A camera is ISO invariant if, when you raise ISO with fixed shutter speed and f-number, deep shadow SNR in the raw data does not change (much). Most cameras are not ISOless for the first few ISO stops but become so thereafter. A good site to check for ISOlessness is sensorgen.info: see when input referred read noise in e- stops changing. I see it does not list the X-T1 but it does show several other Fujifilms. Perhaps one of those uses the same sensor? Or you can do your own SNR tests.
Jack
Hi,
The Fujis listed in the website you gave me are old ones.
I think that’s because recents Fuji’s cameras are not tested by DXO, I think they can’t test the X-Trans sensor which is not a bayer one.
How can I test my camera by myself ?
Is there any application to do this ?
I’m on OSX.
Thanks a lot !
Frankie
You need a tool like RawDigger, Frankie. Put the cap on and take a capture in a dark room at, say, f/22 1/4000s with all NR off at base ISO. Then, keeping f/number and ss the same, take another one at ISO 200, 400, 800, 1600, 3200, 6400.
Take a look at each capture in RawDigger, noting the green channel mean (average) signal and standard deviation – assuming the camera does not clip blacks (I believe the X-T1 does not). Compute SNR = mean/standard deviation, for each ISO. The camera is ISOless when SNR stops changing significantly.
Jack
Hi Jack !
I think it’s totally off my knowledge right now ahah but I’ll try to work on it.
Here’s what I get : https://www.dropbox.com/s/a99aw9b72prirws/X-T1-RAWDIGGER.jpg?dl=0
Hi Frankie, that’s good but you need to untick the ‘Subtract Black Level’ box in RawDigger/Preferences/DataProcessing. Then let’s look at the same data again.
Jack
Hi !
I get a beautiful pink in rawdigger 🙂
Here’s the new data.
https://www.dropbox.com/s/a99aw9b72prirws/X-T1-RAWDIGGER.jpg?dl=0
Thanks a lot for your time Jack !
Excellent. As you raise the ISO the signal should double stop for stop, but you can see that the noise (sigma) doesn’t quite double, making the XT-1 technically ISO variant in the lower ISOs – but only slightly so.
For instance at ISO 200 the green channel sigma is about 1.6 ADUs. At ISO 1600 it should be 1.6*1600/200=12.8 ADUs, but it is only 9.6. So raising the ISO resulted in a slightly better noise performance: log2(12.8/9.6)= 0.3 stops. Not much of an improvement for the 3 stops of highlight headroom given up in the process, still…
At ISO 2000 and above the XT-1 appears to become totally ISOless. Since before that you only gain about 1/3 of a stop by raising ISO, unless you were after every last little bit of noise performance you couldn’t be faulted for treating the XT-1 as ISOless throughout the range.
Jack
Perfect so if I correctly understood, if I want to achieve the BEST quality (considering that I don’t want to loose that 1/3stop of noise) I should raise the ISO in camera at a maximum of ISO 2000 then after this limit I will see no difference between raising my ISO in camera or pushing my Raw in Lightroom.
Thats right ?
Another question, I read a lot of time that Fujifilm is “lying” about the value of their ISO, meaning that ISO 6400 is more equivalent to ISO 3200 for example.
Is there a way to verifying this ?
Note that above I mentioned that the signal ‘should’ double every time you raise ISO. If you do not trust Fujifilm’s ISO to be well behaved you also need to check what happens to mean signal as it is raised.
So you would take a capture of a slightly out of focus uniform subject (e.g. gray card) lit unchangingly and uniformly (ideally stable artificial light in a dark room) at whatever ss and f-number will place the subject’s green channel at around, say, 10000 ADU in RawDigger at ISO 6400 (or whatever maximum ISO you would like to evaluate). Then, while keeping ss and f-number fixed, take a quick sequence of identical captures at ISO 3200,1600,800,400,200,… Note the average green channel values around a 200x200px area in the center of the subject in RawDigger: they should more or less double every time ISO is raised one stop.
If they do, your conclusions above are correct. If they don’t, then we need to continue this chat 🙂
Even if there was no difference between raising ISO in camera or pushing the raw in Lightroom, you better do the former. The manufacturer of your camera has much more knowledge about its sensor than Adobe.
Hi Eric, thanks for your comment. It’s partly a matter of personal taste, but I prefer to use more neutral converters than ACR/LR for large digital amplification (improperly referred to as Exposure compensation) during conversion.
This series of posts makes a very interesting point. I had to think about it quite a bit before I got it because it is so counter-intuitive. To convince myself I did some tests with my Olympus E-M1, taking photographs of the same scene with the same exposure over the whole ISO range. After converting to TIFF in RawDigger to equalize the brightness, I compared the frames side by side at 100 or 200% and found them to be almost indistinguishable. For the most extreme case I did, comparing exposures at ISO 200 and ISO 6400, I could see that when viewing at 100% or more the shadow noise was slightly worse in the ISO 200 frame. Although turning up the ISO to 6400 gave five additional bits of resolution in the A/D conversion, the only perceptible benefit was somewhat finer grained noise in the deep shadows. I also compared the full frame SNR and found it only changed by 0.1 EV.
If I understand this correctly, if there were no downstream noise in the electronics, changing the ISO would have no effect on the image quality. Like most other recent cameras (except Canon) the E-M1 approaches this ideal case because, with multiple parallel A/D converters on the sensor, the downstream read noise (measured) is slightly less than one bit. This has cleared up some of my confusion about what factors affect image quality, by demonstrating that almost the only thing that matters is the amount of light hitting the sensor.
Hi Richard,
Yes, by looking at Bill Claff’s data the E-M1 is not ISOless at least for the first couple of stops above 200 – he shows more than 1 stop difference in the read noise – so there would be some noticeable differences to 6400. What I find amazing, is that when the ISO 200 image is pushed 5 stops to be comparable with the SOOC ISO 6400 one, its visual information is effectively stored in only 7 bits of data , vs 12 bits for the SOOC image. And it’s still hard to tell them apart.
As you say, the average dithering of 0.776 LSB provided by the read noise at ISO 200 is inadequate – I assume that’s what you recognize as quantized noise in the deepest shadows. Contour banding would be possible in such a situation. On the other hand from ISO 800 on up you would be home free, meaning that the data should not be any more quantized in one versus the other.
Good work, thanks for sharing your experience.
Jack
I had not seen the E-M1 in Bill Claff’s data before, just the E-M5. It must be a recent addition. I was pleased to see that the data is almost identical to my own measurements using RawDigger.
I used ImageJ to mask off all but the 5 low bits in the ISO 6400 sample, and found that the result consisted almost entirely of noise plus a tiny bit of highlight detail. I then reread the “Noise, Dynamic Range, and Bit Depth” section of Emil Martinec’s article from 2008, and found that he had explained it all quite well seven years ago. That article is worth revisiting from time to time.
Your articles have given me a better understanding of the difference between data and information.
Yes, Emil Martinec’s pages are THE online reference for noise in digital still cameras, as current as ever. And he doesn’t even list it in his publication history 🙂
Jack I’ve only just come across your blog, and what a find. Thank you for you musings. Although not mathematically in your league I also am try to understand things: photography.grayheron.net
As a Canon user I ‘obviously’ use Magic Lantern. When it comes to exposure strategy I therefore ‘have options’, and have convinced myself that this the following is the best strategy. I would welcome any feedback.
My single-image, i.e. not bracketing, strategy is based on using ISO1600 as the trigger point, i.e. above this ISO my 5D3 behaves as an ISOless camera.
Below this ISO I will use the a base ISO of 100 with ML Raw ETTR to ensure I don’t saturate any highlights. The ML Raw histogram will give me confidence that the shadows, i.e. the DR of the scene is contained.
Assuming the shadows are captured, i.e. lowish DR scene, I will use higher ISOs with ETTR, up to, say, 1600.
Also in this range if the shadows are not ‘containable’ in a single capture I will bring in Dual-ISO, i.e. taking a, say, 100/800 ISO bracket in a single image.
Finally if I feel the need, i.e. stutter speed, to go above an ISO of, say, 1600, I will just do that and leave the ISO at 1600, or even 3200 at a push, i.e. my 5D3 can take that. Thus addressing shadows in post.
I know this is an oldish post, but I would welcome and feedback from you or your readers on what I have written.
Cheers
Garry
Nice site you have there Garry! Sounds like you are going down the right alley. A proper answer would require some more thought but I think this fine Bill Claff chart should be able to give you all the feedback you need.
Jack
Jack
Many thanks for the link.
Cheers
Garry