DxO marks

[joofa]

The underlying physics is that a sensor can distinguish exactly the same colors as the average human eye, if and only if the spectral responses of the sensor can be obtained by a linear combination of the eye cone responses. These conditions are called Luther-Ives conditions, and in practice, these never occur. There are objects that a sensor sees as having certain colors, while the eye sees the same objects differently, and the reverse is also true.

I guess the easy way to visualize the above is as in the graphic below. The recording of color in a device (or human eye) can be modeled as a projection as illustrated below. Different spaces means different "angles" of projection resulting in metamerism and metameric failures. Saying that "Luther-Ives is not satisfied" is equivalent to saying that the two spaces, say sensor and eye, are two different 3D spaces embedded in a larger, higher dimensional space.

[Peter van den Hamer]

I guess the easy way to visualize the above [being metameric failures in color representation] is as in the graphic below.

Cool, Joofa! I recognize the loss of information when going from a spectrum to 3 color channels (RGB).

I am struggling whether this loss of information is equivalent to a lineair projection. So far, I guess you are right: each color channel "sees" the integral of (the light spectrum falling on the sensor x the sensor channel's spectral sensitivity). The spectrum can be approximated as a large number of individual measurements (think: bar chart) that are subject to a lineair transformation aka projection.

Question: when people calibrate colors, they are essentially (non-lineairly) un-distorting a 3 dimensional color space. Why is that needed?
Let's assume, for simplicity, we only want to make the camera see things the exact same way our eyes do (read: let's not get into trying to simulate D50 with tunsten, etc).

Is that because the "projection" from a continous/high-dimensional spectrum to RGB somehow gets distorted? Or is this un-distorting just a hack to compensate for effects of the projection (on "important" colors, like Gretag-Macbeth patches)? In other words, if had a true Luther-Yves conformant sensor, would I still need non-lineair color calibration software (e.g. burried deep down in LR)?

[BJL]

One should not assume that the SOS is used with his camera. The Nikon D800e uses REI (recommended exposure index). My 800e places the metered tone at 14% saturation, 2.82 stops below clipping. This is close to the 12.7% saturation of the old saturation standard and is in line with the DXO measured ISO of 73 for the nominal camera setting of 100. This value is 2.82 stops below saturation allows about 0.34 EV of headroom for the highlights. The saturation standard gives a saturation of 12.7%, allowing 0.5 EV of headroom for the highlights.

The SOS standard allows for some rounding, to accomodate the conventional 1/3 stop increments, and the gap between 12.7% and 14% is within that tolerance. But I agree that in principle it is allowed by CIPA rules for a CIPA member like Nikon to use REI instead of SOS.

How are you measuring this 14%? SOS and REI are defined in terms of output in JPEG or such, not raw files.

P.S. Here are the definitions again, in the original CIPA standards that became ISO standards: http://www.cipa.jp/english/hyoujunka/kikaku/pdf/DC-004_EN.pdf

[joofa]

I am struggling whether this loss of information is equivalent to a lineair projection.

Yes, it is a linear projection.

Question: when people calibrate colors, they are essentially (non-lineairly) un-distorting a 3 dimensional color space. Why is that needed?

Because, by using a nonlinear transformation it is potentially more likely to obtain a mapping between desired and measured responses that satisfies the maximal number of points.

In other words, if had a true Luther-Yves conformant sensor, would I still need non-lineair color calibration software (e.g. burried deep down in LR)?

Luther Ives works only if there are no noise concerns. Therefore, even if Luther Ives is satisfied, noise can make one use nonlinear techniques to get a better response.

[bjanes]

The SOS standard allows for some rounding, to accomodate the conventional 1/3 stop increments, and the gap between 12.7% and 14% is within that tolerance. But I agree that in principle it is allowed by CIPA rules for a CIPA member like Nikon to use REI instead of SOS.

How are you measuring this 14%? SOS and REI are defined in terms of output in JPEG or such, not raw files.

P.S. Here are the definitions again, in the original CIPA standards that became ISO standards: http://www.cipa.jp/english/hyoujunka/kikaku/pdf/DC-004_EN.pdf

BJL,

Thanks for your comments. Since I use raw files for my work, neither the REI or SOS methods are applicable, so I use the saturation standard (as does DXO). Their method is shown here.

I repeated my measurements and used aperture priority for determining the sensor saturation at the nominal light meter reading. This metering mode uses continuously variable shutter speeds and avoids the granularity of the manual shutter speed settings. The sensor saturates at about 15778 in the green channels. At the metered exposure, the raw pixel value in the green channels was 1829, giving a saturation of 12.17%. This allows 0.5 EV headroom for the highlights.

To calculate the ISO, I took shots of my computer screen which has a calibrated luminance of 120 cd/m^2. The sensor saturated with an exposure of 0.85 seconds at f/8. Plugging these values into the DXO formula, I get an ISO of 69, which is close to the DXO measurement of 73.
 
Regards,

Bill
 

[BJL]

Since I use raw files for my work, neither the REI or SOS methods are applicable, so I use the saturation standard (as does DXO). Their method is shown here.

Thanks for the details Bill.

As noted somewhere, when working with raw (and thus linear) data, the saturation measure is essentially equivalent to a mid-tone placement measure akin to SOS, where one for example determines the exposure that places an 18% gray card image at 1/2 stop below 18% of maximum raw level, or about the 12.7% that you mentioned. So there is probably no major difference on that count.

Unfortunately, what none of these approaches measures is something that I would like calibrated, which is exposure index: do the combinations of aperture and shutter speed recommended by the in-camera light meter under given lighting conditions match what you would get from an external light meter set to the same ISO sensitivity as the camera? (Preferably with adjustment for the discrepancy between f-stops and T-stops due to the lens transmission factor as all the ISO and CIPA sandards specify, to keep Ray happy.) Is any camera testing or review site making that sort of check? The REI standard is meant to allow camera makers to approximate this, but with adjustments when using fancier pattern metering systems and algorithms instead of simple center-weighted or spot metering.


P. S. what will or should DxO do if an "ISO-free" sensor comes along, meaning one that does not benefit at from variable analog gain, and a camera with such a sensor uses a fixed analog gain, handling exposure index settings in the digital domain, during raw-to-JPEG conversion, so that raw files reflect variations in EI solely by metadata? In fact, do not some older DMF backs do exactly this?

As far as I can tell, this would lead to the DxO-style raw-based sensitivity readings being the same regardless of the "ISO" setting on the camera. More realistically, it could make sense for analog gain on many modern sensors to top out quite low, between 200 and 800, with all higher EI settings handled digitally. Instead it seems that some cameras do use purely digital gain only beond some level, but apply it to raw files with bit-shifting of the ADC output before recording it in raw files, pointlessly sacrificing highlight headroom.

[Ernst Dinkla]

Update: formula for SMI enclosed.

Hi,

This is what DxO-marks says about SMI: http://www.dxomark.com/index.php/About/In-depth-measurements/Measurements/Color-sensitivity

Best regards
Erik

What is the light source used in the test, the ISO prescription and the one actually used?  Or is the test done over a range of color temperatures or light sources?  I recall something like Canon sensors being more color accurate near 4000K light than at 5000K. Which probably does not give a good SMI quote. DxO  did not test Sigma sensors but it would be interesting to see what their SMI result is.

Luther-Ives condition can not be met by RGB filtered sensors but another eye. Would multi-spectral imaging meet that condition?
There has been the Sony RGBE sensor: RD1http://en.wikipedia.org/wiki/RGBE_filter
For art reproduction multi-spectral imaging is already done.
The HP G4010 desktop scanner uses another approach by scanning twice with different light sources and stacking the samples with adequate algorithms.
The women that have tetrachromacy set another condition too :-)

The Portrait score as a reference to art reproduction like I thought and SMI not counting in "Portrait" scoring seems more or less contradictive. Better keep it at Portrait score if the main ingredient is color noise.

--
Met vriendelijke groet, Ernst

http://www.pigment-print.com/spectralplots/spectrumviz_1.htm
December 2012, 500+ inkjet media white spectral plots.

[hjulenissen]

ErikKaffehr
Yes, Erik, thank you. Its exactly about i worried.
But two-digit index about conformity of only 24 color patches.
(Test target is printed and has much more narrower gamut).

You want something like this:?

[ErikKaffehr]

Hi,


I am somewhat skeptical about DxO and color. The description on their web site is not very complete.

I wrote down a bit about color here, with links to other articles, but I think that color is highly subjective: http://echophoto.dnsalias.net/ekr/index.php/photoarticles/71-mf-digital-myths-or-facts?start=9

Best regards
Erik

What is the light source used in the test, the ISO prescription and the one actually used?  Or is the test done over a range of color temperatures or light sources?  I recall something like Canon sensors being more color accurate near 4000K light than at 5000K. Which probably does not give a good SMI quote. DxO  did not test Sigma sensors but it would be interesting to see what their SMI result is.

Luther-Ives condition can not be met by RGB filtered sensors but another eye. Would multi-spectral imaging meet that condition?
There has been the Sony RGBE sensor: RD1http://en.wikipedia.org/wiki/RGBE_filter
For art reproduction multi-spectral imaging is already done.
The HP G4010 desktop scanner uses another approach by scanning twice with different light sources and stacking the samples with adequate algorithms.
The women that have tetrachromacy set another condition too :-)

The Portrait score as a reference to art reproduction like I thought and SMI not counting in "Portrait" scoring seems more or less contradictive. Better keep it at Portrait score if the main ingredient is color noise.

--
Met vriendelijke groet, Ernst

http://www.pigment-print.com/spectralplots/spectrumviz_1.htm
December 2012, 500+ inkjet media white spectral plots.

[qwz]

hjulenissen
Yes, but MaxMax.com uses JPEG - and this method completely wrong.

[qwz]

And about DXOs new results.

А99 - 1555 Low-Light ISO
RX1 - 2534 Low-Light ISO (pretty similar to the same sensor in D600)

As we know, a99 mirror takes around 1/4 of light.
How exactly DXO measured such big difference???

[hjulenissen]

hjulenissen
Yes, but MaxMax.com uses JPEG - and this method completely wrong.

Do you have a reference for this?

I did not know that they used JPEG (I had a look at their website and the text seemed to suggest that they used raw, though they did use imprecise language). I can not imagine how one would use JPEG for this, as WB and color correction should make any measurement strange and meaningless.

-h

[dreed]

And about DXOs new results.

А99 - 1555 Low-Light ISO
RX1 - 2534 Low-Light ISO (pretty similar to the same sensor in D600)

As we know, a99 mirror takes around 1/4 of light.
How exactly DXO measured such big difference???

Maybe the impact of the mirror on low-light ISO performance is not linear?

[BJL]

And about DXOs new results.

А99 - 1555 Low-Light ISO
RX1 - 2534 Low-Light ISO (pretty similar to the same sensor in D600)

As we know, a99 mirror takes around 1/4 of light.
How exactly DXO measured such big difference???

Because DxO is not measuring the signal in photoelectrons counted by the sensor: it is looking at raw output levels produced at the various ISO exposure index settings on the cameras, after analog gain and analog-to-digital conversion. Thus, a camera maker's decision to apply a greater or lesser amount of analog gain at a given ISO setting will give a higher or lower DxO sensitivity "score", even from the same sensor tested with the same exposure H in lux-seconds (e.g same intensity of sensor illumination L in lux and same exposure time t in seconds; H = L*t, ISO defined exposure index EI = 10/H.)

Let me explain once again that DxO is _not_ measuring a pure sensor (photosite) characteristic, but a "processing pipe-line" characteristic including the effects of gain applied after the signal in photoelectrons is moved from the photo-sites.

By the way, my analysis of the DxO results for some different cameras with the same sensor confirm that the horizontal scales on its SNR and DR curves do not compare sensors at equal exposure in lux-seconds, so are misleading in comparing performance in low light conditions. You get far closer to comparison at equal exposure (i. e. at equal exposure index) if you move each dot on those curves horizontally back to the values 100, 200, 400 etc. indicated by the cameras' ISO sensitivity settings. That is, compare at the ISO exposure index values used by the camera, not the DxO raw file level saturation measurements.

[Vladimirovich]

By the way, my analysis of the DxO results for some different cameras with the same sensor confirm that the horizontal scales on its SNR and DR curves do not compare sensors at equal exposure in lux-seconds, so are misleading in comparing performance in low light conditions. You get far closer to comparison at equal exposure (i. e. at equal exposure index) if you move each dot on those curves horizontally back to the values 100, 200, 400 etc. indicated by the cameras' ISO sensitivity settings. That is, compare at the ISO exposure index values used by the camera, not the DxO raw file level saturation measurements.

can you show us that analysis please ? you probably had 2 raw files and we can see raw histogram from rawdigger to illustrate the saturation, right ?

[Ray]

You get far closer to comparison at equal exposure (i. e. at equal exposure index) if you move each dot on those curves horizontally back to the values 100, 200, 400 etc. indicated by the cameras' ISO sensitivity settings. That is, compare at the ISO exposure index values used by the camera, not the DxO raw file level saturation measurements.

I'm not sure what point you are trying to make here, BJL, unless your point is that the noise performance of two cameras used at equal shutter speed can vary to an even greater extent than it would if different shutter speeds were used to compensate for different ISO sensitivies.

If two cameras have the same fundamental performance in terms of SNR and DR etc, but one of the cameras has a third of a stop lower sensitivity, as measured by DXO, then that camera with the lower sensitivity will require a third of a stop slower shutter speed to achieve an equally correct or full exposure, exhibiting the equal SNR and DR that it is potentially capable of.

However, graphs do have a vertical and horizontal axis, so it should be no problem to guess what the difference in SNR and DR will be when two cameras are used at the same shutter speed. This gives one an appreciation of just how significant in practice, the different 'real' sensitivies of sensors may be.

For example, the DXO graphs for SNR, comparing the Sony RX-1 with the A99, indicate that using the same shutter speed with the RX-1 and A99, at at their base ISOs of 100 for the RX-1(actually 81), and ISO 50 for the A99 (actually 48), would produce at least a 2dB worse SNR on the A99. One can guess this value simply by moving one's eye vertically down from the best reading for the RX-1 (at ISO 81) to an imagined ISO 81 on the A99 graph.

However, if one uses whatever shutter speed is required to produce a full exposure with each camera at its base ISO, the differences are less. SNR at 18% is a negligible 0.2dB down for the A99, but DR is 1/3rd of a stop down, which is getting close to significant.

These results are roughly consistent with the fact that the A99 has a fixed, semi-transparent mirror which unavoidably redirects a small portion of the light. Consequently, at all manufacturer-nominated ISOs, one would expect the A99 to require a longer exposure to achieve, hopefully, the same or at least nearly the same performance.

The interesting point is, even with the longer exposure, the noise and DR performance of the A99 is not quite as good as one might expect. The third of a stop lower DR performance at the different base ISOs, is perhaps not significant, but the 2/3rds of a stop difference at the manufacturer-nominated ISOs of 800, is significant. In other words, after giving the A99 a good 1/3rd of a stop more exposure than the RX-1 at ISO 800, the RX-1 still retains a 2/3rds of a stop DR advantage.

If one were to use the same shutter speeds with both cameras at the manufacturer-nominated ISOs of 800 (and same FL of lens at same aperture and same T/stop of course), the RX-1 would have a full stop DR advantage, and as much as 3dB SNR-at-18% advantage.

Such differences can be of practical significance when shooting in full manual mode, selecting a specific shutter speed required to freeze movement, and a specific F/stop for DoF and/or sharpness purposes. The RX-1 in these circumstances, according to DXO's measurements, should produce approximately (or as much as) a whole stop better DR and about 2/3rds of a stop better SNR, assuming lenses with the same T/stop are used. That's quite a significant improvement.

Tonal Range and Color Sensitivity will also be noticeably better in the RX-1 at equal shutter speeds and F/stops.

There's also another issue which still puzzles me a bit, which Bart has not yet clarified    .

The RX-1 has a fixed lens, and therefore any ISO sensitivity measurements by DXO must include the lens which will have a specific T/stop factor unavoidably included in the ISO sensitivity results. The T/stop factor is a transmission loss due to the opacity of the lens elements, which effectively reduces the size of the aperture regards exposure requirements. The lens on the RX-1 at F2 might effectively be F2.5 regards shutter speed requirements. I'm not aware of any lens that has a T/stop rating which is numerically smaller than the actual F/stop rating. That would imply a light transmission gain, instead of a loss.

When DXO test the A99 for ISO sensitivity, they do not use a lens. However, the photographer who uses the A99 has to use a lens, and that lens will inevitably introduce another factor which effectively reduces the sensitivity of the sensor, by some degree, whether by 0.1 of a stop, O.25 of a stop, 0.67 of a stop, or sometimes even by a full stop.

So those figures I've extracted from the DXO graphs (and not from my arse, in case you are wondering), are probably very conservative. The best case scenario of a 0.2dB loss in SNR, and a 0.32EV loss in DR, comparing base ISOs at the appropriate shutter speed for an ETTR exposure, could only be reached if the lens used with the A99 had a T/stop equal to its F/stop, which is perhaps unrealistic.

Agreed, Bart? 

[BJL]

Ray,

    If at the same ISO setting and corresponding exposure level (aperture-shutter speed combination or to be more precise, lux-seconds), two cameras give different level placements in raw files, I see no evidence of a need to adjust exposure level. One can instead achieve correct levels in the final displayed output (JPEG or whatever) by appropriate raw conversion. For example, if the midtone placement is at the Ssat "half stop down" level, then then a default raw conversion with flat tone curve and normal contrast will need to apply a half stop adjustment, multiplying the linear levels by 1.4 at some stage. If instead the camera maker has decided to provide a full stop of highlight headroom, the factor is 2 instead of 1.4; if 1 1/2 stops, then 2.8.

And indeed, raw conversion software generally knows about the different midtone placement in different raw files (because there is no industry standard recommending or requiring any particular raw level placement, and as DxO has shows there is not a hint of uniformity in this respect) and so default conversions seem to give a reasonable midtone placement in JPEG (around 116-118).

At higher exposure index levels (say 800 and up for the EM5) where noise is the greatest concern, cameras are amplifying the dark noise up to well above the quantization noise level of the ADC, so quantization noise is insignificant and different raw placements have no significant effect on the S/N ratios of the data recorded in the raw file. Thus after the appropriate level scaling in conversion from raw, these different choices about raw level placement have no significant effect on noise levels in the final displayed image.

On the other hand, if for whatever reason your goal is to have midtones placed in raw files exactly a half stop below 18% of maximum level, you still do not need to change exposure levels (aperture/shutter speed). Taking the simplified example of a one stop difference (a camera than places metered levels 1 1/2 stop below 18% of maximum raw level), you could
- use the same exposure levels (for example with EI of ISO 800)
- set the camera's ISO one stop higher (1600 in this example)
With in-camera metering, this corresponds to setting a +1 exposure compensation;
with an external meter, it corresponds to setting that meter to an EI of ISO 800.


Do not worry; this does _not_ give SNR at the level indicated by the DxO measurements of 18% SNR for the higher ISO=1600, because the one stop "overexposure" is placing metered light levels twice as high as those used to get the 18% SNR measurement for ISO 1600 setting, and that increase in illumination improves SNR by a factor of 1.4, or 3dB higher: almost exactly the same as the values at ISO 800 setting! No surprise, because SNR 18% is dominated by photon shot noise (until you reach very high exposure index) and so is determined mainly by the count of photons detected by the photosites, very little effected by in-camera noise sources, and for a particular sensor, the photon count is determined by the exposure (lux-sec) delivered to the sensor: that is, by the chosen exposure index, not the ISO setting in the camera or the DxO SSat value.


As an example, I compared two recent cameras with the same sensor (and I have verified equal quantum efficiency): the Olympus E-M5 and E-PL5: see E-M5 and E-PL5 at DxO

Analyzing the SNR graphs, it is clear that the values on the SNR and DR curves at equal camera ISO settings are in fact at equal exposure levels (shown by near-equal 18% SNR values) so the only difference is that as ISO setting in increased, the gain applied between photosites and raw levels differs, increasingly higher for the E-PL5 than for the E-M5. DxO uses a horizontal axis of DxO SSat values, pushing the dots on the E-PL5 curve to the right and therefore moving the SNR and DR curves "higher", but if instead a graph used exposure levels (equal aperture/shutter speed combinations under equal lighting conditions), the curves would basically agree, with some small fluctuations. This agreement shows that if one proceeds to expose at exposure index equal to camera ISO setting, you get the same SNR at sensor exposure levels all the way down to seven stops below those for 18% at base ISO speed (which is what the 18% SNR value for ISO 25600 setting measures, indirectly.)

That is, in this comparison, variations in the raw level placements used with the same sensor has no significant effect on noise levels in the images produced with equal exposure (lux-seconds), contrary to what is suggested by the DxO SNR graphs, and in the difference in DxO "sport" score between these two cameras. The lower raw placement of the E-M5 does not at all hurt its IQ at equal exposure level.

[BJL]

Ray,

    I basically agree with what you say about using DxO data to assess base-ISO sensitivity. The DxO measurements at a camera's minimum ISO setting (and thus minimum gain between photosites and raw level) is a useful measure of the maximum exposure level (minimum exposure index) than can be used and still have a standard "safe" amount of highlight headroom before some highlight photosites get saturated. (Note the fun fact that the ISO 50 setting on the RX1 has exactly the same DxO SSat of 81 as its ISO 100 setting, so the latter is probably already sending full well signal to maximum raw level, and ISO 50 is a one stop "overexposure" with the same gain from photsites to raw levels as ISO 100.)

Also, with the DxO data showing that the A99 has about 1dB worse SNR at equal exposure levels, and so needs about 1/3 stop more exposure to get equal SNR, then this indicates a true sensitivity difference of about 1/3 stop, and I agree that this is probably attributable to the partial mirror of the A99. More generally, DxO SNR 18% graphs can be used to get relevant comparisons of noise levels at equal exposure level, if one reads the graphs correctly.

Indeed, reading where the three horizontal lines on the DxO SNR 18% graphs hit the curve gives measures akin the the ISO 12232 noise-based sensitivities, except that DxO has chosen to mark SNR levels of 38:1, 30:1 and 20:1 rather than the more traditional 40:1 ("excellent image") and 10:1 ("acceptable image") mentioned in ISO 12232 and in this Kodak document: http://www.kodak.com/global/plugins/acrobat/en/business/ISS/supportdocs/ISOMeasurements.pdf

To get noise-based upper limits on usable EI levels, choose a SNR level like the green 38:1 line, read across till it hits the SNR curve, find the dot closest to that value, and read of the exposure level that achieved this SNR. To be more precise, find the two dots on ether side and interpolate the exposure levels. Of course the best estimate of exposure level available from DxO is the manufacturer's ISO, not the DxO measurement of raw level placement, for reasons that I have explained repeatedly.


Apart from this special case where a partial mirror reduces the effective sensitivity of the sensor (maybe DxO could break the A99's mirror off to give us true raw sensor performance measures!) have you found examples where different cameras using the same sensor give significantly different SNR 18% at equal ISO settings?

[Vladimirovich]

The lower raw placement of the E-M5 does not at all hurt its IQ at equal exposure level.

now imagine that you do raw+jpg shooting... with the approach when camera undersaturates you either have a good raw (and too bright JPG) or good jpg (and undersaturated raw)... that is not a problem for JPG shooters (if they like in camera JPG from a particular camera) or for raw shooters as they saturate as they want (at the expenses of usable preview).

[BJL]

now imagine that you do raw+jpg shooting... with the approach when camera undersaturates you either have a good raw (and too bright JPG) or good jpg (and undersaturated raw).

I do not need to imagine, as I do use RAW+JPEG with the E-M5, and do not see any problem with either raw or JPEG files. The in-camera JPEGs have appropriate levels (not to bright or too dark), and there is no detectable problem with the raw files, with the default conversions by various raw convertors giving JPEGs similar to the in-camera ones default. And as I have tried to explain and illustrate, the lower raw level placement does no measurable harm to noise levels for the E-M5, which is near the extreme for low raw level placement.

But perhaps by "undersaturated raw" you simply mean a raw file with the metered exposure levels placed at less than 12.8% of the maximum level (three stops below maximum), so that highlights can be more than three stops above metered level before they suffer clipping (which at higher than base ISO speed would be due to excessive amplification not over-full wells). My recurring point, as illustrated with the above comparison of the E-M5 to the E-PL5, is that within the range of placements used by various cameras, this seems to cause no detectable harm to the final image quality; it just requires multiplying the raw levels by a factor different from 1.4 at some stage of the raw conversion process.

So what exactly is the disadvantage you see in placing the metered exposure level at less than 12.8%?

I suspect that some people are still confusing two things:

• maximizing signal levels (photon counts) in the photosites while not clipping highlights, by giving the maximum safe exposure level at the photosites, meaning minimum safe exposure index: "Expose To The Right".
  
• maximizing the numerical raw levels, by using the maximum amplification of the photosite output that does not amplify any highlight pixels up to clipping level, beyond maxim ADU output level: I iwl call this "Amplify To The Right".

These are quite different at higher ISO settings: for example, at exposure level of ISO 1600, midtone photosites might have 2% of full-well capacity, while even if highlights range up to four stops brighter, the brightest are still at only 32% of capacity. But if raw conversion then places the midtones at 12.8%, those highlights are blown: in fact the signal from any photosite receiving more than 16% of capacity gets "clipped" in the amplification and ADC process. In that situation, the photosites are all under-saturated, but the ADC output is "over-saturated".