DR daydreaming (channeling my procrastination towards interesting stuff)

[Samuel H]

This Friday I had a big shoot and I should be editing but I'm mentally tired and doing all sorts of things just so I don't sit down and edit. Like: I just spent three hours running DR tests to see what Sony and Samsung and Nikon could give us if they wanted.


This is a summary of the results.


* I set the a7S to slog2 (Flaat) and shot some video and RAW stills tests. The RAW stills ahve around 1.25 stops more DR than the video footage. In particular, video loses one stop of DR in the shadows, in part because the internal codec kills some texture detail, and in part because slog2 is not as aggressive as it could be. The part that is due to the codec -no idea how big it is- should be solved by recording external 4K ProRes HQ.


* Then I set it to "no PP" and shot the RAW stills tests again, at ISO 3200 and ISO 800. Slog2 has two extra stops in the highlights wrt usual picture profiles, so my guess here was that slog2 ISO 3200 is just ISO 800 underexposed by two stops, then remapped with a different gamma curve. The results support this. The no-PP 800 stills look a lot like the slog2 3200 stills, whereas the no-PP 3200 stills have a stop and a half less DR (also, they have clearly less DR than slog2 video; therefore ISO must be, electronically, a different thing when using slog2 or no-PP).


* Just for fun, I ran the test again with "no PP" and ISO 100. This has 1.25 stops more DR than both no-PP 800 and slog2 3200 RAW stills. This tells me that the a7S could record at ISO 400 or 800 using a profile that's even flatter than slog2, and it would have between 1.5 and 2 stops more DR in video mode than it does right now. Even if the a7S has more DR than most of its competitors, Sony is holding back on the DR front.

The a7S could shoot slog2 at ISO 400 (elecronicallally-same-as-no-PP-ISO 100). It would have the same DR as it has now, maybe with a bit less noise in the shadows.
With the same settings (elecronicallally-same-as-no-PP-ISO 100) and slog3, it would be ISO 1600 and it would have around two stops better DR than it currently has (1.5 because of slog3, 0.5 because of less noise in shadows). I would overexpose it by one or two stops, and treat it as ISO 800 or 400. If it is kept at 8-bit, on the upper side of the curve it goes from 29 values per stop of light to 20 values per stop of light.




* I ran the tests with my NEX-5N too, at ISO 100, with RAW stills. It came out 1.5 stops short of the a7S at ISO 100, and half a stop short of a7S at no-PP 800 (or slog2 3200). According to dxomark the difference should be between 0.3 and 0.5 stops. I suspect the dxomark method is hurting the a7S relative to others for some reason, just like I think it's hurting the NX1. You have to decide many things when measuring DR, it's far from straightforward or purely objective. For example, I decided highlights are only usable if they are correctly colored: washed out highlights didn't cut it on my test, skin goes funny when that happens to it. But for clouds this wouldn't be an issue, and counting desaturated highlights as usable would probably change the relative results. The usual DR target test is monochrome, and this includes the one dxomark uses, so this could be the source of the difference. I designed the test to fit *my* needs.


* The problem with the unused power of the a7S is that Sony is unlikely to unleash it: it has to protect its higher end cameras. But Samsung doesn't have that "problem". And according to dxomark the NX1 at ISO 100 matches the DR of the a7S at ISO 100. So, those 1.5 to 2 stops more DR can come from Samsung if they release a wider gamma curve that doesn't clip any information coming out of the sensor. I know this would lead to banding issues and whatnot, but I'd rather fight banding than clipped highlights.


* A hack that records RAW on Nikon DSLRs would be really huge. According to dxomark, the D810 has 1.7 stops more DR than the a7S shooting RAW stills at ISO 100. Maybe the dxomark procedure is hurting the a7S relatively more than others, and maybe that'd be just 0.7 stops with my methodology. But also: RAW doesn't have these 8-bit-related issues the NX1 would have, nor the compression issues of the a7S slog2 video. a7S RAW stills at ISO 100 already have 2.5 stops more DR than a7S slog2 video. With a D810 shooting RAW video, I believe we'd have 3 extra stops over the a7S slog2 video. Mouthwatering.

 

[DLD]

Various Sony rumors have any of A7R MK II, A9 or A99 MK II being announced later this month. Let's see what they decide to tweak first. A thin body will probably have the 5-Axis stabilization but not 4K. A DSLR style (translucent mirror) big body might have both.

 

[Samuel H]

a7RII and a7SII with IBIS and redesigned a7II body must coming for sure in the next few months, but since the a7S was already video-oriented and it's clear Sony was holding back (even if it is great), I woulndn't hope for a lot of progress. Probably just IBIS and the new body. Internal 4K would be great but they may keep it out until the III version, next year.

Also: given my results, and given the fact that smaller sensors are faster to read out, I think an a6100 with slog2 could be just as good as the a7S in terms of sharpness and DR. It would make for a great B-cam to mount on my N4000 (I finally managed to get my a7S plus vintage primes working there, but it works a lot better with a lighter load).

 

[photochris]

Also: given my results, and given the fact that smaller sensors are faster to read out, I think an a6100 with slog2 could be just as good as the a7S in terms of sharpness and DR. It would make for a great B-cam to mount on my N4000 (I finally managed to get my a7S plus vintage primes working there, but it works a lot better with a lighter load).

I'd like to see this as well. Sometimes a small camera is just so handy (not a fan of the LX100) and the Sony 10-18 is a great UWA lens. I briefly used the A6000 with the 10-18 as my UWA since its so much smaller and about the same price used relative to a good 16-35 for a FF camera - and you get fewer lens changes, extra body and so on.

 

[paulcurtis]

* I set the a7S to slog2 (Flaat) and shot some video and RAW stills tests. The RAW stills ahve around 1.25 stops more DR than the video footage. In particular, video loses one stop of DR in the shadows, in part because the internal codec kills some texture detail, and in part because slog2 is not as aggressive as it could be. The part that is due to the codec -no idea how big it is- should be solved by recording external 4K ProRes HQ.

Lets explore this some more.

I have side by sided Slog and Cine to see whether the 8 bit container is limiting range. But instead of looking at RGB i started off looking at the actual YUV raw data from quicktime. I am going through an external recorder so those 8 bits are being packed into a 10 bit container. My concern was that because of the raised blacks in slog then i would see quantisation errors in that recording.

I discovered that there was no appreciable difference between cine and slog beyond the obvious gamma encoding.

I also discovered that by doing the YUV -> RGB conversion myself i could extract a bit more highlight.

The Slog curve this camera uses, especially over HDMI is NOT the same as any other Slog curve, the blacks are lower and if you try to use a normal Slog LUT you will clip the shadows.

The movie output DR isn't as high as stills but some of that i believe is that there's no highlight reconstruction going on with the movie. It is what it is because white balance is baked in. So highlight clipping, even though better that most, is still harsh. I'd be considering ways to try an get some highlight reconstruction going on with this data in YUV but not cracked that yet. Don't forget that in stills and linear RAW there is so much data at the top end that these things can make a big difference.

As you say the shadows are compromised too, the splotchy noise in the shadows mystifies me - the stills are clean so where is this noise coming from?

The choice of default ISO is weird too. Side by side if you process slog2 and cine1 to the same exposure the noise is the same. So i don't believe 3200 is actually 3200 or 500 in cine1 is actually 500. I suspect we don't *know* what the actual ISO of the camera is, instead we have these arbitrary indications. I believe that 500 in cine1 is identical to 3200 in slog2.

I wish for RAW output over HDMI, that would be awesome. Or pre white balanced data.

I feel the colorspaces for the movies are compromised more than anything else. I can achieve skin tones in RAW that i've simply not been able to match in Slog and it's really annoying.

cheers
Paul

 

[DLD]

Various Sony rumors have any of A7R MK II, A9 or A99 MK II being announced later this month. Let's see what they decide to tweak first. A thin body will probably have the 5-Axis stabilization but not 4K. A DSLR style (translucent mirror) big body might have both.

Well, a thin body is in with 4K and 5-Axis stabilization.

And a bunch of pixels too.

And rumors were generally proven correct.

 

[Ebrahim Saadawi]

It's strange why they offer an HDR mode (S-LOG) but only make it limited to be used at sensor ISOs with Lower DR, not the highest (100). Just strange.

It's a given fact that in all sensors, all of them, from Canon, Nikon, Samsung, Sony, Toshiba, Aptina, have the highest DR at the base ISO of 100ish (in some cinema sensors 650/800) then starts dropping with applying gain. (with the weird exception of the C300/II C-Log claiming perfectly consistent 15 stops of DR from 650 to 25600 ISO, which I don't understand as gain noise must decrease the shadows DR by a large amount, anyhow)

so why not take that 100 ISO image and apply a LOG gamma squeezing all the DR if I am going for HDR video?

Maybe because they found this is the maximum DR to be contained within 8bit therefore decided to lock it at minimum 800 gain?

Or that it would need a newly designed LOG gamma and codec to contain higher than that and that's just not practical.

The sensors are capable of higher DR than we're getting on the A7s, D810, A7r, 5D, even 550D, and the only way to get it all it get a raw video signal at base ISO (the only way to exploit the extremist sensor potential) or an aggressive LOG gamma with a fat codec.

I always say it's marketing and ''stick it to the man'' but it turns out to be a technical limitation that I am just 100% clueless about, like internal 4K on A7s/5D.

 

[Ebrahim Saadawi]

If DXO DR measurements are viable then for example, an ancient Nikon D5100 (500D rival) at base ISO has higher DR than the A7s, and with RAW/Log to harrness it should be alexa league.

Not to mention the D7200, the second ranking highest DR camera in the entire data base (FF and MF)

The old A7 supposedly also should be able to achieve 14 stops, higher than A7s, if given raw/log

A7s

And the A7s, should be able to reach 13.5ish DR if LOG/Raw shot at 100 ISO. While clearly, at 3200 ISO which S-LOG2 claims to be, the maximum sensor performance is somewhere at 10 stops, not much higher than a 5DIII at 3200ISO, so theoritically that's all we should be getting is S-LOG2 (and 9-8 stops going 5000-6400 ISO and up) which is clearly not true, so it reinforces your assumption it's not 3200 ISO gain, but somewhere around 400-800 where the maximum DR is 12.5-ish, and S-Log squeezes it all. So they can squeeze the more DR at 100 ISO on the A7 (14)/A7s (13.5) but the only limitation I can think of is that they would need to design a new tailored LOG gamma for the specific sensor, and/or they find an 8bit container can only carry 800 ISO 12-ish stops so they limit it to that, or they just don't want their 1000$ A7 with S-LOG to have 14 stops of DR like their F65.

 

[Samuel H]

There are two things playing here. One is a technical thing related to slog2 and similar gamma curves, and another is a Sony decision I also don't understand, which may or may not be based on technical things.

The technical thing: since slog2 and similar gamma curves have more headroom in the highlights than standard gamma curves (around two stops in the case of slog2, 3.5 in slog3), if you leave your sensor electronically the same, you have to shoot less light at your subject, which means you have to treat it as higher ISO. If the a7S had slog2 ISO 400, the sensor would be doing exaclty the same job as when shooting Neutral with ISO 100. So, ISO has to be higher because the curve demands more headroom in the highlights.

But it doesn't have to be as high as Sony made it on the a7S. And that's the Sony decision I don't fully understand. From my tests,

The a7S could shoot slog2 at ISO 400 (elecronicallally-same-as-no-PP-ISO 100). It would have the same DR as it has now, maybe with a bit less noise in the shadows.
With the same settings (elecronicallally-same-as-no-PP-ISO 100) and slog3, it would be ISO 1600 and it would have around two stops better DR than it currently has (1.5 because of slog3, 0.5 because of less noise in shadows). I would overexpose it by one or two stops, and treat it as ISO 800 or 400. If it is kept at 8-bit, on the upper side of the curve it goes from 29 values per stop of light to 20 values per stop of light.

They probably decided to keep slog3 out of the a7S because it is 8-bit and 20 values per stop of light is going to show bad banding. I would love it to be an option, though: much easier to fix that than blown-out highlights. But they get to choose what they sell.
And they raised the minimum ISO to 3200... for some reason that may or may not be technical. I would think it's just a way to protect their more expensive cameras, but also if you look at the very low-level tests, the a7S does have a change in the way the sensor works from ISO 1600 onwards: link link. Maybe slog2 just needs one stop more range than their other modes and for some reason they really really really want to be using that mode for slog2, even if it has less DR than sticking to electronically-same-as-ISO-100.

 

[DLD]

... The sensors are capable of higher DR than we're getting on the A7s, D810, A7r, 5D, even 550D, and the only way to get it all it get a raw video signal at base ISO (the only way to exploit the extremist sensor potential) or an aggressive LOG gamma with a fat codec.

The DxO numbers are for Raw stills and one can't get Raw out of a small stills/video hybrid cameras currently due to - mainly - price considerations. C300 MKII (and Arri Alexa) managed to do it with the dual-readout. Both camera prices have reflected it (Alexa from 2010'ish and C300 MK II since a month from now). When you think about the fact that FS700 + Odyssey Q7+ can get Raw for ~ $6,500 already and BMD is promising to deliver it for $5,000 in-body URSA MIni, then one suspects that video products are more or less capable of matching their sensor maximum capabilities even at a budget level.

As to why Log is set at higher ISO's I assume it's also due to the processing requirements for the data pouring in lower ISO's.