It would sell alot

Basically yes – you would push the SNR to a higher value thus gaining room (=DR) towards SNR = 2 and SNR = 1. Or in other words, pushing the noise floor down. However this SQRT relation is the theoretical maximum if the noise is completely uncorrelated and if downsampling is done “properly”. In my experience, you gain less than that in reality. We typically talk about 0.7 – 1 stop gain at e.g. SNR = 2 if 4 pixels are downsampled into 1.

So just to be clear, the expected SNR when downscaling is going to be:

Measured_CineD_SNR * sqrt(number of photosites / number of pixels at the downsampled res)? So then if I’m lucky, there will be a row in the Imatest screenshot that corresponds to this SNR and tell me how many stops are above this point. Is my understanding correct?

as Guest, you wrote above that that a camera’s dynamic range (DR) is maximum saturation to sensitivity threshold, not noise threshold. Well, what determines the sensitivity threshold? Usually that minimum value _is_ the noise floor, that’s how DR is determined or defined. A signal has to be greater than this to be recognised as a signal. (Put simply, not going into statistical noise analysis.) In systems where there is a minimum threshold/gate or say “forward bias” that is above any meaningful noise, this would then change the definition to the lowest signal threshold, and not noise, but I would regard that as specialised cases. For systems such as camera sensors where the photoreceptors are designed to be as sensitive as possible, ready to capture photon signals from their noise floor, the noise floor is part of the DR determination. Though sometimes the noise floor can be truncated too, to clean an image.

(I should mention too that for linear DR would be at the minimum saturation point, not maximum saturation point.)

Noise can be reduced by various approaches, and one approach is downsampling. In this discussion case, spatial downsampling is the concept. Spatial downsampling will reduce the statistical variance of the noise value, thus lowering noise power. As the noise floor is lower, the dynamic range will be increased. The signal will be spatially downsampled too, so there may be loss of the minimal signal too (compromising DR gains depending on the signal across the downsampled pixels), and of course the pay off is lower resolution. There will be dynamic range gains, though only fully realised for lower spatial frequency projections on the neighbouring photoreceptors.

This whole discussion started with Matt Williams question to the merit of comparing pixel noise to a standardised final resolution. He declared “I’m not an expert when it comes to testing stuff like this”, but the question is sound nether-the-less. Downsampling to 4K would seem like a good contemporary standard for noise comparison purposes, depending how the DR values are determined in the first place.

Is there anything fundamentally flawed in my reckoning? I felt compelled to respond when you wrote that dynamic range is not a ratio to the noise threshold, as in an engineering sense, that is the very definition in understanding what dynamic range is actually about, if time is taken to ponder what the minimum value really is.

Putting everything in a strict term of E(ngineering)DR is arguably the most STEMlord and Debatelord behavior I’ve ever seen. If so, why don’t you specify the temperature for dark noise and the different read noise in different readout modes?

There is a certain limit when we are discussing the theoretical definition and the general perception of it.

And this is where your STEM brain fails, it is a fact that downscaling doesn’t introduce new information, but it does filter out a certain amount of useless information. And no one would think that denoise can magically increase any information: it is pure lunacy and daydreaming.

However, from a practical perspective, it makes the useful information more identifiable, which is the CORE of any information theory. That is also 1.why noise threshold matters (and again, if someone thinks that a SNR=0.000001 is usable then I would just smile and move on) and 2.a latitude test matters because people adjust exposure in post and see if they can get a good result, instead of looking at a Xyla21 chart, counting photons, or spend time online spouting uninformative sxxt.

Just wrote above bunch of useful stuff, your comment shows you missed it and you just digress. And I’m not worrying here about dynamic range, but preventing delusions and having to re-educate clients or colleagues expecting increasing DR in post because they read it on the internet.

And “aaah” – yes.

Camera’s full dynamic range is the range between maximum saturation capacity and sensitivity threshold. Range between maximum capture point and minimum. Not noise threshold. The latter can be affected by post processing so if you take that as a reference for full dynamic range you may imagine you increase dynamic range in post. Which is a fantasy.

You have full camera DR visible in those numbers. That largest number.
Lower ones are due to different S/N criteria to establish usability of full captured dynamic range.

DR didn’t go up in post, it was already there in capture but interpreted differently, you just got back closer to original larger figure by changing S/N ratio through downscaling.

Hey Gunther, ppl are agressive because they feel entitled to get their journalism for free. Pay no attention to this, if the man wants articles with scientific depth, perhaps he should subscribe to a scientific paper and pay a fee for content.