Sony A7sII Initial Impressions

[paulcurtis]

RAW stills should be identical, no hurry on that front.

An improvement using slog2+sgamut is something I didn't expect. Very welcome anyway, but I'd go straight to slog3+sgamut3.cine anyway: banding is easier to fix and less bothering than clipping (denoise, downsample, add grain, banding should be gone) (also: there are many reports of image issues on the a7S that in the end turned up being a result of the LUTs used or the software used to apply them).

All this is looking pretty good, thanks again to everybody sharing their results.

Samuel,

Here's a side by side of A7s and A7sII both at 3200. Both Slog2/SGamut. You can see clear differences for the reds. The two 50mm lenses flare differently hence exposure differences, the A7sII had a voightlander 1.5 and the A7s had the contax zeiss 1.4. Needless to say the A7sII looks like how the scene was.

So i think this may well evidence that the A7s had some seriously screwed up SGamut science. To be honest so far i'm not seeing that much difference between SGamut and the .cine version.



In addition here's a screen showing the YUV comparison between Slog2 and Slog3 both recorded off the A7sII on the O7Q. You can clearly see the encoding difference. The waveform is just displaying the Luma, the Y part. The Slog2 is making more of the space and given the same clipping point you can see that Slog3 could hold a lot more range than this camera can deliver.

Now the big question is internally how is this encoded, especially in the XAVC files. Is it as simple as 0 to 255 integer values for the luma channel? Or is the encoding done differently? In other words is Slog3 only using code values 30 to 210 out of a possible 0 to 255? If this is the case then Slog3 will not be able to represent tonal values as well as Slog2 (which in turn is only borderline good enough compared to the cines).

This is made more difficult with the ProRes from the O7Q because these are 10 bit files, so 8 bit stuffed into 10 bit files but in order to read them via QT the pixel format is r4fl (that's 32bit YUV) but when you ask QT for that format it always dithers the image. If you ask for 16 bit YUV the range is different and it dithers. Only in 8 bit will it not dither but then you're not actually asking for what's in the file. Does that make sense?

So if we cannot work out the encoding scheme then we need practical image tests to eyeball.

https://www.dropbox.com/s/okxbj74xrvxig5u/yuv.jpg?dl=0



cheers
Paul

 

[adaptation]

As usual, great stuff paul. They most definitely resolved the issues of S-gamut a7s. I will be curious to see how well slo3 sgamut3 grades with all that information squeezed in there. I could see the same issue we had before where the colors stayed washed and faded no matter how much saturation we put in - the only solution was to crank in camera saturation to 20+ . Keep up the good work.

Samuel,

Here's a side by side of A7s and A7sII both at 3200. Both Slog2/SGamut. You can see clear differences for the reds. The two 50mm lenses flare differently hence exposure differences, the A7sII had a voightlander 1.5 and the A7s had the contax zeiss 1.4. Needless to say the A7sII looks like how the scene was.

So i think this may well evidence that the A7s had some seriously screwed up SGamut science. To be honest so far i'm not seeing that much difference between SGamut and the .cine version.

View attachment 106376

In addition here's a screen showing the YUV comparison between Slog2 and Slog3 both recorded off the A7sII on the O7Q. You can clearly see the encoding difference. The waveform is just displaying the Luma, the Y part. The Slog2 is making more of the space and given the same clipping point you can see that Slog3 could hold a lot more range than this camera can deliver.

Now the big question is internally how is this encoded, especially in the XAVC files. Is it as simple as 0 to 255 integer values for the luma channel? Or is the encoding done differently? In other words is Slog3 only using code values 30 to 210 out of a possible 0 to 255? If this is the case then Slog3 will not be able to represent tonal values as well as Slog2 (which in turn is only borderline good enough compared to the cines).

This is made more difficult with the ProRes from the O7Q because these are 10 bit files, so 8 bit stuffed into 10 bit files but in order to read them via QT the pixel format is r4fl (that's 32bit YUV) but when you ask QT for that format it always dithers the image. If you ask for 16 bit YUV the range is different and it dithers. Only in 8 bit will it not dither but then you're not actually asking for what's in the file. Does that make sense?

So if we cannot work out the encoding scheme then we need practical image tests to eyeball.

https://www.dropbox.com/s/okxbj74xrvxig5u/yuv.jpg?dl=0

View attachment 106377

cheers
Paul

 

[boorye]

Received the A7sii and put it through some tests... So far, so good. One thing to note is that the 1080p 120p is not fixed to FF mode but actually fixed to APS-C/crop mode. Rolling shutter in 4K is terrible. Any sort of fast camera movement and your screen turns into jello. It gets much, much better in crop mode. I think I have myself a new B cam, fellas!

 

[paulcurtis]

Received the A7sii and put it through some tests... So far, so good. One thing to note is that the 1080p 120p is not fixed to FF mode but actually fixed to APS-C/crop mode. Rolling shutter in 4K is terrible. Any sort of fast camera movement and your screen turns into jello. It gets much, much better in crop mode. I think I have myself a new B cam, fellas!

120p is a centre crop of the full 4K sensor, it's apparently a crop of around 2.2x. By doing this means the read out is much faster than the full frame (or even an APS-C crop) therefore rolling shutter is negligible in that mode. However the recording quality is pretty poor.

I would like to have seen a centre crop 1080p at normal frame rates too - to minimise RS for certain circumstances yet retain better compression.

RS in 4K is no worse than any other dSLR style solution IMHO.

cheers
Paul

 

[nyvz]

HD from FF? Yes it downscales from 4K to 1080p and it's damn good. I've some other questions and tests i'd like to check first. One being the idea that is it better to record internal HD at 50mbits or UHD at 100mbits and scale under all scenarios. So on a static or slow shot clearly the UHD downscale is better. But on a high detail and moving shot does the UHD break up faster than the 50mbit of HD, bearing in mind UHD is 4 times the area for only twice the data rate.

cheers
Paul

Yeah I've been surprised how flimsy 100mbps 4k on my a7rii seems even compared to avchd on other cameras which seems odd considering I figured it would just be like 4 avchd frames stitched together... So far the 4k doesn't seem to handle noise very well so it's very likely in low lighting that 50mbs hd would produce somewhat better results than 100mbps 4k. I wonder if it's a limitation of the 4k encoder in the a7-series.

 

[paulcurtis]

Just a interim post, so you can see what goes on behind the scenes with compression.

The screenshot below shows the U channel from the YUV. YUV is an encoding scheme that separates the luma from colour. The UV channels are a colour difference scheme, the UV channels are always very compressed. The values in the UV channels are very narrow. What i've done is boost the contrast/black and white points so you can see the kind of data that is used to rebuild colour. Shocking isn't it. And remember this is the ProRes 422 HQ version! Now you know where all the colour noise comes from.

These YUV values are run through an expression to give RGB. Normally this is an operation that QuickTime will do behind the scenes however there are more ways to convert and sometimes you can get some more range out of the source YUV by doing it yourself (i've managed another half stop but it's variable)

I'm curious whether this is the output from the camera or the ProRes compression.

I'm not sure whether i can capture uncompressed via the O7Q (i know i can do dpx but that's RGB only)

The levels of noise in Slog2 vs Slog3 can look different but really if you match the exposure for a certain area then they seem pretty similar.

Now can we get RAW please Sony... (these are the kinds of things that separate pro cameras from prosumer.)



cheers
Paul

 

[Samuel H]

View attachment 106376

Awesome stuff, paulcurtis, thanks a lot. It's great when first users are really knowledgeable.

Looking at this sample... WOW, that's a huge difference! But it doesn't seem like a fixed-sgamut kind of difference to me: it looks like a much lower noise floor coming out of the sensor. Remember how the a7S had much lower DR than the a7 or the a7R in RAW stills in the dxomark test? Looking at this, my first guess would be that they did a metal spin on the processor design, keeping the basic design but moving a bit some routes in one of the layers to fix crosstalk issues or something like that. We'll know if this is the case when the dxomark test comes out.

So that's the good news, now for the bad news... by definition, slog3 has 1.3 stops more DR in the highlights than slog2 (see the curves here), but it seems the a7S II is not fillling that extra space: there's a hard clip at exactly the same point. If that's the case, the DR improvement will be a lot smaller than I expected (it will go from 2 stops, 1.3 from the highlights and 0.66 in the shadows, to just 0.66 in the shadows).
But also this test shot has the white clip coming too fast for this to be measured, it's a test for something different. The easiest way to check the white clipping point this is to point a strong light against a wall, close and at an angle, and set the camera so there's clipping in part of the wall but not all. Shoot with both cameras, and see if the clipping point is closer to the light on the II with slog3.

 

[paulcurtis]

Awesome stuff, paulcurtis, thanks a lot. It's great when first users are really knowledgeable.

Looking at this sample... WOW, that's a huge difference! But it doesn't seem like a fixed-sgamut kind of difference to me: it looks like a much lower noise floor coming out of the sensor. Remember how the a7S had much lower DR than the a7 or the a7R in RAW stills in the dxomark test? Looking at this, my first guess would be that they did a metal spin on the processor design, keeping the basic design but moving a bit some routes in one of the layers to fix crosstalk issues or something like that. We'll know if this is the case when the dxomark test comes out.

So that's the good news, now for the bad news... by definition, slog3 has 1.3 stops more DR in the highlights than slog2 (see the curves here), but it seems the a7S II is not fillling that extra space: there's a hard clip at exactly the same point. If that's the case, the DR improvement will be a lot smaller than I expected (it will go from 2 stops, 1.3 from the highlights and 0.66 in the shadows, to just 0.66 in the shadows).
But also this test shot has the white clip coming too fast for this to be measured, it's a test for something different. The easiest way to check the white clipping point this is to point a strong light against a wall, close and at an angle, and set the camera so there's clipping in part of the wall but not all. Shoot with both cameras, and see if the clipping point is closer to the light on the II with slog3.

Thanks Samuel,

The purpose of that image is just really to show the difference in colour with both being SGamut, perhaps it's not so obvious where i've got the wipe between the sides but the reds are totally different between the two. And red has *always* been the issue.

I'm going to do some more controlled noise tests this evening i hope, and also shoot some back lit wedges for range.

I'm not sure there is any improvement in DR overall. Certainly the noise is better. I don't know whether they're just exposing up and taking down a stop in camera now, i don't think so though but it's always a possibility.

The original DR of the A7s didn't really fill Slog2 either, and certainly not Slog3 but even so i was still surprised how low it clipped. So before i really put it out there i want to double check what i'm doing and looking for.

There has always been a mismatch between the stills and movie that i never understood. Why can't we have the range shown in raw?

cheers
Paul

 

[paulcurtis]

Of course now i'm eyeing up the Leica CL...! 10 bit 4k? Mmm...

cheers
Paul

 

[paulcurtis]

Here are some initial dynamic range tests. What i am showing right now is the raw YUV data, the luma data from the externally recorded ProRes files. These are UHD recorded but down sampled to HD, including the A7s. I've not denoised them. This is the raw luma data in the file. Next i will process them correctly via their various gamma curves but this provides a nice way to look at the basic output.

Observations. Well the noise on the A7sII is clearly better. Compare the Slog2 from the A7s with the same ISO below and the noise is very different. Under that is the 1600 ISO version, perhaps a tiny bit better than 3200 but very little in it.

What i'd advise is to download the PNG and take that into photoshop and look at the levels so you can crush and examine highlights. On every strip the second to the left wedge is the one that should just clip.

I have always maintained that the A7s is around 11 stops and that can be seen clearly here. (each wedge is 1/3rd stop)

I'd say the A7sII is basically the same, but that last stop is more usable. Whereas the temptation in the A7s is to crush it.

Slog3 is the disappointment here. Clipping never goes to white, it clips at 90% (out of 109% on HDMI). It also appears more noisy in the shadows too. The Slog3 curve can hold 15 stops (about 1.5 more than Slog2), so the A7s series can pour 11, maybe 11.5 stops into that 15 stop bucket. Therefore that bucket will never get filled to the top - hence clipping is quite low. I suppose that's just the way it is. So i'm a bit disappointed that this camera isn't offering more range. I'm pretty sure the sensor range is there (in RAW) but not output perhaps.

Conversely cine4 for comparison is practically noiseless at 400 ISO, it does retain highlight detail to the second wedge like the others but the highlight region is super crushed. It is there though.

So far i'm not sure why anyone would use Slog3. I need to keep looking and testing. I think Slog3 is an easier format to work with in Post and perhaps that's why it is here.

Here's a link to the file https://www.dropbox.com/s/1pcangi9efpipnf/wedges.png?dl=0



I'll now process the wedges using their respective gamma curves.

I'll also be looking at comparing the colour spaces next. Although i can tell you that SGamut3.cine and SGamut3 just tweak the primaries, i cannot see any changes to the gamma (which is to be expected).

Cheers
Paul

 

[ntsan]

My A7S II test results: http://www.eoshd.com/comments/topic/18425-a7s-ii-in-stock-at-samys-la-quick-review/

SLog3 has some banding issues which can be fixed using SLog2 + SGamut3.cine. Stock PP3 worked well in low/mixed light (see link above).

100Mbs is as other have stated way too low for 4K. Still not bad compared to 100Mbps GH4 4K (though GH4 has the edge with less macroblock artifacts).

The A7S II is a worthy upgrade from the A7S, IBIS is really good. It does overheat, though only so far on long takes (testing- I'd never shoot that long- see link for more details).

The new Sony FE 24-240 is a nice all-around full-frame lens, similar to the 18-200 on the A7S (APS-C mode) in versatility (except close focus).

The A7S II color is a really nice improvement, but not as good as Canon 5D3 RAW for skin tones.

Thanks for the test! So it still can overheat, bummer

 

[Erik Naso]

I just ran a continuous record test in 4K. Recorded for the full 30 minutes in a 73 degree room with no overheating.

 

[jcs]

I just ran a continuous record test in 4K. Recorded for the full 30 minutes in a 73 degree room with no overheating.

So did mine. It overheated after restarting after 29:59, about 10 minutes in (74F).

 

[jcs]

So far i'm not sure why anyone would use Slog3. I need to keep looking and testing. I think Slog3 is an easier format to work with in Post and perhaps that's why it is here.
l

Great test- thanks! So far I've found SLog2 gamma with SGamut3.cine to be a decent combo for color and noise. Need to test SLog2 with Pro color (what I used the most on the A7S).

 

[jcs]

Favorite settings so far:

High dynamic range: Slog2 + SGamut3.cine, +16 saturation, -7 sharpness
Indoors: Cine1 + SGamut3.cine, -7 sharpness.

120fps + SpeedBooster: looks really good, less artifacts than the FS700 (didn't see any objectionable artifacts when viewed at normal zoom).

Shooting RAW stills- skintones are looking really good, a big improvement over the A7S. Need to test side-by-side next to the 5D3 to compare skintone color.

 

[paulcurtis]

Great test- thanks! So far I've found SLog2 gamma with SGamut3.cine to be a decent combo for color and noise. Need to test SLog2 with Pro color (what I used the most on the A7S).

With regard to colourspaces. All they are is a set of RGB primaries that specify what 'red' , for example, actually means. Ultimately you'd be viewing on a device which also has a set of primaries, usually sRGB.

So in theory it shouldn't matter what colorspace you use so long as

a) That space is big enough to encompass all the colours the camera can produce
b) You convert colourspaces in post correctly

I don't believe this camera can produce anything outside Rec709, but that needs testing.

The difference with the A7sII so far is that the whole colour work camera side is improved, better skin in RAW stills is a good indication because i had to jump through hoops before to get decent skin out of stills.

Previously we were using the colorspaces to minimise the bad color but there was always an issue of workflow stepping in, in that i believe a lot of the issues were workflow based.

The downside about colorspaces is that if you choose one that is so wide and fill just a little bit of it then in that 8 bit container you are actually loosing colour definition.

I posted an image of what the UV channels look like and they look, well, awful but that's the nature of 8 bit YUV. The range of data in the UV channels is really small, so anything we do to make that smaller will have a more negative effect.

So what i'm hoping to do is to compare by looking at these UV channels, more quantifiable. But also eyeball the colour overall.

It's worth pointing out that SGamut and SGamut3 have the same primaries. SGamut3.cine has slightly different ones. In fact SGamut3.cine is a *smaller* space than the others. Which is why it may well be the best option for us.

I have a hunch that you've found the best combination, Slog2/SGamut3.cine but by adding saturation you're making it non standard and you're no longer able to interpret that footage as SGamut3.cine. Conversely perhaps by boosting saturation here you're making better use of the limited colour bandwidth. So that's not clear either

cheers
Paul

 

[Samuel H]

Thanks Erik Naso and jcs!
No overheating on 30 min takes is a nice positive surprise, I was expecting just 5 minutes.

Thanks paulcurtis!
That DR test is really useful. A big letdown, but absolutely crucial information. My biggest reason for getting an a7S II was the extra DR from slog3, which the camera CAN deliver (the a7S already could). An improvement of 2 stops might be worth the $$$$ for me, just 0.66 stops is definitely not.

The problem with sgamut is that it doesn't work as you described, because it's not linear. If you get people in front of the camera and ramp the shutter speed up and down, you'll see their skin tone going green or magenta depending on how bright it is on the viewfinder. It's like having a color look applied to your image right out of the camera. Sgamut3.cine is linear by design, it does what you're describing.

With what we know so far, I would also go for slog2+sgamut3.cine, with standard saturation if you're going to use a LUT (e.g. from LUTcalc, going for LC709A, which makes it Alexa-like in a very specific sense, link, link, link), with extra saturation if you're doing all post corrections manually (that's how I've been shooting with the a7S, because yes, I think it makes better use fo the limited color bandwidth).

 

[stokstad]

The combination of offering proffesional gammas but not good enough quality (bitrate, colordepth, sampling etc) to use them properly is really strange. To use Slog3 to preserve 13-14 stops of DR in a compressed format seems like a good idea. To use Slog3 for LUT-workflows seems like a good idea (IMO the only viable idea). The A7-series seems to instead offer these gammas and gamuts for better colour-rendition and highlight roll-off, but you wont get the DR those curves were made for, and the 8-bit makes LUTs problematic. They should probably make a 12 stop LOG for all their 8-bit cameras like canon did, because the workflows are becoming ridiculous (recording to external in 4K just to get a decent 1080P).

Im not saying these are bad cameras, they seem very good indeed, but its like they all tick more boxes than any other cameras but at the same time they all miss one or two really important boxes. Market separation i guess

 

[Samuel H]

Market separation for sure. This thing can make much better use of slog3, keeping the white clipping point at the same place where it is on slog2 is a marketing decision. They want me to think "slog3, I'll get it!" but then they also want people to spend twice as much to get the FS5 with ACTUAL SLOG3. I feel like when a car manufacturer puts a high-end brand on the surname of a lesser model (or when Hasselblad puts its name on Sony cameras, for that matter).

 

[paulcurtis]

Thanks Erik Naso and jcs!
That DR test is really useful. A big letdown, but absolutely crucial information. My biggest reason for getting an a7S II was the extra DR from slog3, which the camera CAN deliver (the a7S already could). An improvement of 2 stops might be worth the $$$$ for me, just 0.66 stops is definitely not.

I know what you mean but i believe that the shadow stops are more useful now, so there is an advantage in that. I'm very curious why though. I do wonder whether the RAW dynamic range is a result of post processing in camera, which is why we don't see it in movies. Things like highlight recovery and the way it's all matrixed from RAW

The problem with sgamut is that it doesn't work as you described, because it's not linear. If you get people in front of the camera and ramp the shutter speed up and down, you'll see their skin tone going green or magenta depending on how bright it is on the viewfinder. It's like having a color look applied to your image right out of the camera. Sgamut3.cine is linear by design, it does what you're describing.

Well i wonder whether that effect you're describing with green and magenta changing with exposure is actually the camera doing colour wrong in the first place and not SGamut as such? Sony publish the colour primaries and they define a colour space, they don't publish exposure based primaries. I do remember something about white balance being baked in though which always made no sense to me. What makes you think SGamut is non linear and why? Curious to learn.

The combination of offering proffesional gammas but not good enough quality (bitrate, colordepth, sampling etc) to use them properly is really strange. To use Slog3 to preserve 13-14 stops of DR in a compressed format seems like a good idea. To use Slog3 for LUT-workflows seems like a good idea (IMO the only viable idea). The A7-series seems to instead offer these gammas and gamuts for better colour-rendition and highlight roll-off, but you wont get the DR those curves were made for, and the 8-bit makes LUTs problematic. They should probably make a 12 stop LOG for all their 8-bit cameras like canon did, because the workflows are becoming ridiculous (recording to external in 4K just to get a decent 1080P).

Im not saying these are bad cameras, they seem very good indeed, but its like they all tick more boxes than any other cameras but at the same time they all miss one or two really important boxes. Market separation i guess

I suspect they're in there to make the footage play well with other footage as a B Cam. Slog3/cine3 is such a good standard at the moment and you can just slot this footage straight into an edit or pipeline. Just because *we're* obsessive nuts that try to maximise the camera doesn't mean others care so much!

You can add as big a colour bucket as you like but if this camera (and in fact pretty much all cameras) can't fill it then it's more of a future proof standard.

cheers
Paul