Sony a7S III Lab Test – Does it Live Up to the Hype?

October 6th, 2020
Sony a7S III Lab Test - Does it Live Up to the Hype?

A production version of the long anticipated new Sony a7S III has finally arrived at the CineD headquarter for our lab test. How does it fare in comparison to its peers? Read on …

EDIT: According to requests we get, will test the ISO 12,800 value soon and update the article with our findings. Thank you for your patience.

Before we start, a little personal note from myself – when the predecessor, the Sony a7S II hit the shelves in Vienna on October 2015, I was rushing to the stores and bought it immediately as this was the camera to buy for full frame videography. I was very happy with it for more than 3 years, and end of 2018 I sold it as I was thinking a successor should come soon … well, it took quite a bit longer than that as we all know.

Now in case you missed our real-world review (of a pre-production Sony a7S III), watch and read it here.

The 10-bit Sony – good enough in 2020?

Now we have October 2020 and the Sony a7S III is in my hands – and I am asking myself if I should rush to the stores again? After all, the camera landscape has changed and a lot of competing 10-bit full frame models from other manufacturers are readily available (Canon, Panasonic, Nikon, and the offerings from ZCAM just to name a few). Plus, in the meantime I bought a BMPCC6K – and I am super happy with it …

Nevertheless, running the Sony a7S III through the lab was a task that I was looking forward to very much.

All the tests below were performed in SLOG3, SGamut3.Cine with the internal XAVC S-I all intra 4:2:2 10bit codec.

Sony a7S III: Best rolling shutter value ever measured

As usual we are using our 300Hz strobe light for this as explained previously. And voilá: the Sony a7S III trumps (sorry for using that word) – rewind – the Sony a7S III has the best full frame rolling shutter readout value ever measured: a whopping 8.7 ms for UHD (3840×2160) in 25 frames per second.

This value is really a huge step forward, so far the best value was from the Canon R5 in full frame readout having 15.5 ms (in 4K DCI 17:9 mode which has 7% less picture height). For comparison, the recently measured Panasonic S5 clocks in at 21ms in UHD.

Now, quite interestingly with 8.7ms a maximum of ~115 frames per second would be possible. Looking at 100fps in UHD mode, the same 8.7ms are measured, hence up to 100fps nothing changes in the readout – very fascinating.

But then, switching the Sony a7S III to 120 frames per second a slight crop is applied and the rolling shutter value reading goes to 7.7ms – now we know why the image is cropped slightly to make the 120 frames per second possible.

After this superb start, lets have a look at the dynamic range.

Dynamic Range of the Sony a7S III at ISO 640 SLOG3 SGamut3.Cine

Shooting the DSC Labs Xyla 21 stepchart (see a writeup of our test procedures here) the following waveform plot is obtained:

Waveform plot of the Xyla 21 stepchart of the Sony a7S III at ISO640 in SLOG3 / SGamut3.Cine in UHD 25p mode: around 13 stops are visible above the noise floor.

Around 13 stops can be identified above the noise floor. However, what becomes obvious is a very clean noise floor – it looks like a lot of in-camera noise reduction is happening which cannot be turned off. This was to be expected from the low-light king which has always been Sony’s a7S line.

IMATEST calculates the signal to noise threshold value of 2 at 12.4 stops – a very good result but this has to be put into perspective.

IMATEST dynamic range reading of the Sony a7S III at ISO 640 in SLOG3 / SGamut3.Cine in UHD 25p mode: a solid 12.4 stops are calculated.

Looking at the middle chart of the three IMATEST graphs above the in – camera noise reduction becomes obvious – above the light blue line (signal to noise (SNR) threshold of 1) no further stops are visible towards the dark (left) end. This is always a sign of noise treatment in one way or another, and the risk is that the image doesn’t look organic any more.

Comparing this to recently tested cameras like the Panasonic S5 (please have a look again here) you can see that the S5 has additional stops visible in the noise floor and still shows 12.1 stops at the SNR = 2.

Hence, a solid result at SNR = 2 for the Sony a7S III but don’t expect much more information to be retrieved from the overly clean noise floor.

Dynamic Range of the Sony a7S III at ISO 16000 SLOG3 SGamut3.Cine

As my colleague Johnnie noticed in his review of the Sony a7S III, there seems to be a second native ISO at 16000 where the image is very clean.

Hence I decided to look at the dynamic range at ISO 16000 as well. Again, looking at the waveform below 12 stops plus a faint 13th stop are visible. Quite amazing, considering the high ISO value.

Waveform plot of the Xyla 21 stepchart of the Sony a7S III at ISO 16000 in SLOG3 / SGamut3.Cine in UHD 25p mode: around 12 stops are visible above the noise floor.
IMATEST dynamic range reading of the Sony a7S III at ISO 16000 in SLOG3 / SGamut3.Cine in UHD 25p mode: a solid 11.9 stops are calculated.

IMATEST calculates 11.9 stops at SNR = 2, which is really amazing for ISO 16000. So you lose very little going from ISO 640 to ISO 16000.

As Johnnie called it in his review, the Sony a7S III is a “night vision” device.

Latitude (underexposure) Test of the Sony a7S III

As part of every lab test we perform the latitude test. Latitude is the capability of a camera to retain image detail and colors when over- or underexposed. Very revealing, as not only the sensor capabilities are displayed but also the internal codec capabilities – how well is shadow information retained.

For our base exposure scene we adjust the strength of the fixed studio light such that at an f-stop of 4 my colleague Johnnie’s face is exposed at a maximum of 60% Luma value in the waveform display for 25 fps, 360° shutter angle for ISO 640 (of course, Johnnie’s position and relative distance to the grey background is kept the same for all cameras).

Then, we successively underexpose the scene but decreasing the shutter angle to 180°, 90°, 45°, 22.5° and finally 11.25° – hence underexposing by 1, 2, 3, 4 and finally 5 stops.

In DaVinci Resolve Studio (using the latest version 16.2.7) we push the image back to the zero base exposure by the adjusting lift, gamma and gain sliders and comparing the waveforms.

Jumping ahead to the 3 stops underexposed and pushed back scene, we get the following image:

Sony a7S III three stops underexposed and brought back in DaVinci Resolve Studio. Noise shows up in the darker areas, but still usable.

For 3 stops of underexposure, pushed back the image is still quite good, however in the lower right hand side in the dark areas nasty blocks of chroma noise are raising their ugly head. But the image is still usable.

Sony a7S III four stops underexposed and brought back in DaVinci Resolve Studio. Larger patches of ugly chroma noise are appearing – the limit is reached.

At four stops of underexposure, brought back to zero the limit of usability is reached. While the colors and image detail are still largely there, ugly blocky chroma noise artefacts are appearing (look at the shadow area at the lower right side) which I find impossible to remove via noise reduction.

Here is a try in DaVinci using 3 frames temporal, threshold 5 and spatial threshold 1 – for me the limit before the image detail is lost and everything becomes plasticky:

Sony a7S III four stops underexposed and brought back in DaVinci Resolve Studio plus additional noise reduction. The larger patches of ugly chroma noise are impossible to remove with noise reduction without impacting the overall image detail.

Just for reference here is the 5 stops underexposed image, which is completely unusable, with or without noise reduction:

Sony a7S III five stops underexposed and brought back in DaVinci Resolve Studio. The image falls apart.
Sony a7S III five stops underexposed and brought back in DaVinci Resolve Studio. Due to the blocky patches of chroma noise applying noise reduction doesn’t help.

Unfortunately, a similar situation like we encountered with the Sony FX9 in the lab test is happening: it seems that the internal codec XAVC S-I (4:2:2 10bit) with an average bit rate of 250Mbit/s for 25 frames per second is not capable to retain shadow detail without showing nasty chroma noise artefacts. Those large patches of blocky chroma noise cannot be removed effectively by noise reduction.

Summary

The new Sony a7S III shows a mixed bag in the lab test. It shows by far the best rolling shutter result that we ever tested for a full frame camera – 8.7ms. Amazing.

In the dynamic range test at ISO 640 it has a solid 12.4 stops at SNR = 2, but not much more is there – maybe half a stop. It looks like Sony has a lot of internal noise reduction processing going on to squeeze out this result. Which for most users is fine – a great out of the box result! At ISO 16000 the results are really impressive: 11.9 stops (SNR = 2) are measured.

A Panasonic S5 for example shows 12.1 stops (SNR = 2) but a 13th and 14th stop are still there – similar to the Canon R5 in RAW filming mode. And the Sony FX9 has a 13th, 14th and even a hint of a 15th stop buried in the noise floor, although it shows only 11.5 stops at SNR = 2 – thats the reason why we always give you the full result that you can put things into perspective.

In the underexposure test, the Sony a7S III is hindered by strange, large patches of chroma noise artefacts – if you don’t want to have those the limit is reached at 3 stops of undereposure – rather similar to the Sony FX9 if you use the internal All-I codec.

For comparison the recently tested Panasonic S5 actually looks surprisingly good until 4 stops of underexposure, you could get away even with the 5 stops underexposed image. Mainly because the noise is very finely dispersed such that you can very effectively apply noise reduction in post production (and yes, please turn off the lens compensation for V-Log as our reader Cliff Totten has pointed out).

In any case, I want to remind everyone that the lab test is just one piece of the puzzle when comparing / making a judgement on cameras – overall the Sony a7S III packs a huge list of superb features in a very compact body.

Times are good for us filmmakers.

Have you had some experience already with the new Sony a7S III? Please let us know in the comments.

81
Leave a Reply

guest
1
1
- 3
Filter:
all
Sort by:
latest
Player008
Player008
Guest
14 days ago

You should use the Exposure Correction plugin for DaVinci Resolve to adjust your exposure settings in your tests: https://bit.ly/3nOaW3V

This will be a far more accurate solution for exposure adjustment and measurements instead of the manual methods you are currently using.

Last edited 14 days ago by Player008
Kris Smith
Kris Smith
Member
October 11th, 2020

I’m confused about your comments on 100FPS. Everything I’ve read says there IS the 1.1 crop at 100p, not just 120, but your tests say the crop is only in 120?

Ilya Zakharevich
Ilya Zakharevich
Guest
October 11th, 2020

BTW, it seems that you are measuring the performance of camera+lens, but you never mention the lens. Or maybe you know that the photon noise from the diffused light is negligible comparing to the sensor noise?

(With many sensors of today, the sensor noise may be ∼1.2e₋ at the high gain. Already 0.5e₋/pixel of stray light would increase the total noise by 16%!)

Hmm, this was the noise for 14-bit readout modes. The sensor noise should be more for 12-bit readout. Mutatis mutandis!

Last edited 17 days ago by Ilya Zakharevich
Ian Hunter
Member
October 9th, 2020

If it’s not good enough for you, go create your own camera company.

Last edited 18 days ago by Ian Hunter
Dano Motley
Dano Motley
Guest
October 9th, 2020
Reply to  Ian Hunter

?

Dano Motley
Dano Motley
Member
October 9th, 2020

You lost me at the Trump remark. Sad that you let politics seep into your well done review.

Nick
Nick
Guest
October 9th, 2020
Reply to  Dano Motley

Get over yourself! You made a bad choice by following a poor human being. It’s ok Dano to realize that.

Dano Motley
Dano Motley
Guest
October 9th, 2020
Reply to  Nick

Nick you assume who I will vote for and really dont know…do you? I am just sick and tired of politics. It would be better to leave it out of a camera review.
Nick I sure as hell did’nt ask for your opinion…

Dano Motley
Dano Motley
Guest
October 9th, 2020
Reply to  Gunther Machu

No problem at all. Just tired of politics and the covid flu in everything. Great review though…thanks.

Werner Botha
Werner Botha
Member
October 9th, 2020

Why not test HLG as well?

Raw Shooter
Raw Shooter
Guest
October 8th, 2020

Philip Bloom says that at high iso’s you really need to shoot raw or the images are kinda junky. But with raw they’re recoverable. I can’t see where you say you did that. Maybe I missed it.

Patrick Grossien
Patrick Grossien
Member
October 8th, 2020

Thank you so much Gunther, for the gained insight with all your tests and comparisons.
Since this is going to become even more relevant the more cameras you tested, I’d love to see a comparison tool from you guys where we could compare the test results between cameras with links to the in-depth reviews. Maybe it’s already in the works, but if not – pretty please! With a cherry on top 😉

Ilya Zakharevich
Ilya Zakharevich
Guest
October 8th, 2020

A lot of thanks for your important investigation. However, I must note that one of the main components:

 > “above the light blue line (signal to noise (SNR) threshold of 1) no further stops are visible towards the dark (left) end. This is always a sign of noise treatment in one way or another, and the risk is that the image doesn’t look organic any more.”

is only a result of misunderstanding how the Imatest works. FACT: the Imatest algorithm for determining the input-related noise is
(undocumented [?] and) broken. It seems that it cannot deal with very long almost-horizontal pedestals on the left of the S-curve. And, to add insult to injury, they do not subtract the (calculated) black level from the S-curve — which might have worked around this defect.

Observe the dotted green line on the top graph. It SEEMS that the middle graph is just the bottom graph divided by the slope of this green graph. So when the slope is close to 0, you see that “spike” on the middle curve. CONCLUSION: near the minimum of the green curve, the middle graph makes no sense.

Observe that NOTHING SPECIAL happens in this zone on the bottom curve. There is no “special noise reduction” happening in this zone. What you observe is just an artifact of (the BAD, green) polynomial approximation to the S-curve.

This also shows that in presence of such pedestals, one should take the L-dynamic range reported by Imatest with a GREAT grain of salt. It may be strongly under-estimated.

Last edited 20 days ago by Ilya Zakharevich
Ilya Zakharevich
Ilya Zakharevich
Guest
October 8th, 2020
Reply to  Gunther Machu

Note that I specifically said “input-related” noise. So far, I did not find any problem with their calculation of the “pixel-related” noise. (However, given how “mathematically hapless” they seem to be, I reserve my doubts. ;-)

 > “ (i.e. is an artefact of the flat pedestal).”

Nope, it has very little to do with flat pedestal. (Except that in absence of the pedestal, this bug is MUCH less pronounced.)

The reason for the bug is that to calculate the “input-related” noise, one needs to calculate “the local γ”. However, instead of sound algos to calculate “the local γ”, Imatest uses a polynomial approximation — where it is obvious that a polynomial approximation has no chance to work.

Ilya Zakharevich
Ilya Zakharevich
Guest
October 8th, 2020
Reply to  Gunther Machu

OK, let us go into math details:

(A) polynomial approximations work with (VERY roughly speaking) curves with “‘regularly’ changing data”. When you take log WITHOUT subtracting the black level, you get an “artificial” pedestal — which, essentially, “follows a different law than the rest of the graph”. (There the curve is “flat”, while in the rest of the curve, it is linear — such curve “switching from one law to another one”, cannot be easily approximated by polynomials.)

(B) If they subtracted the black level, the pedestal on the log-log curve would disappear. As a result, even their borken algorithm (using a polynomial approximation) would have a good chance to get to no fake division-by-0.

(C) If P is a good approximation to the log-log S-curve (with subtracted black point!), then Input␣noise = Pixel␣Noise/P’. If P is a bad approximation, then P’ gets “fake” zeros (the easy-to-observe “artifact” minima on the dotted green curve) where this formula “bombs out”.

(D) Thinking about this more, I’m getting more and more afraid that they may COMPLETELY IGNORE the black point, so just divide by P’ for the BIASED log-log curve! If so, their data is becomes more and more bogus the closer one gets to the black point. (Here I suspect errors of order of a couple of STEPS in the “L”-dynamic range!)

Last edited 20 days ago by Ilya Zakharevich
Ilya Zakharevich
Ilya Zakharevich
Guest
October 8th, 2020

I’m afraid that (D) has a chance to hold! I eyeballed the graphs for α7sIII, S5 and R5, and they have something very similar to what (D) predicts!

(1) The prediction: the input-related noise is overestimated by a factor 1+B/s, here B is the black level, and s is the (linear) sensor’s readout. This means that the “middle” graph would have a corner near s=B, and the slope on the left is 1 more than the slope on the right.

(2) Assuming 1/γ=2.5 (when corrected for black level) for shadows (which fits the linear part of “the top” graphs for α7sIII, S5 and R5), the level s=B is achieved γ log₁₀2≈0.12 steps above the pedestal (in units of the vertical axis on the “middle” graph). This matches density about -3.4 for α7sIII and R5; likewise, it is about -3.7 for S5.

(3) I did not try to data-process it, but there ARE such corners VISIBLE on the middle graphs! (The only mismatch is that the corner seems to be closer to -3 for R5.)

This should be taken with a grain of salt — since the eye can believe anything — but it is a reason for further investigations!

Last edited 20 days ago by Ilya Zakharevich
Ilya Zakharevich
Ilya Zakharevich
Guest
October 9th, 2020
Reply to  Gunther Machu

Very cool idea! Thanks! Below is my musing about possible enhancements to your “trick”.

About your choice of the “modified” black level. By default the graph shows about 8–9 patches on the pedestal. I presume that after change 9→1, this becomes 5–6 patches? This is still too many for the polynomial approximation to work… (I expect that degree=5 they use should be able to cope with 2–3 patches.)

I never worked with dVR, but I presume that “luma” matches the percentage in the digital range (0…2¹⁶−1) reported by Imatest. (So luma=9 matches 9%, which matches log₁₀=‒1.05 of the “top” graph). Given that the log₁₀=-1 is crossed between 7th and 8th patches, this suggests that the “TRUE” black level is 0.01% below the minimal patch.

However, using this “averaged” black level would clip the noise. So it is better to decrease the black level, say, about 1.7σ of the pixel noise. (As usual, Imatest muddies the report by using strange formatting, but) I think that the pixel noise is about 0.22% in the dark patches. This means that converting black level to luma≈0.37 should not clip the input too much. — And this may gain another 1.5 stops in the region of “linearity”.

Ilya Zakharevich
Ilya Zakharevich
Guest
October 9th, 2020

Just to put things in context: my experiments with 7-year-old 12-bit sensor (one which does not use RAW noise reduction) show that it is very linear IN THE RANGE OF 19.5 bit (sic!). Basically, this means that noise reduction has a chance to extract details up to 0.005 of the digital unit of RAW output. (This means S/N ratio of about 0.01 MIGHT be useful — although in very restricted contexts.)

THIS is why above I focus on a possibility to get the input to the format which Imatest has a chance to handle without triggering bugs — so it has a chance to report a meaningful “middle” graph above the S/N=1 line.

Anyway, what you report rejects my (D)-guess! (Although I cannot invent any algorithm which would have this “divide by 0¯bug, while keeping the “L” value so precise…)

Ilya Zakharevich
Ilya Zakharevich
Guest
October 9th, 2020

IIRC Imatest allows a text report of its data on the patches: the average recorded signal per patch, and RMS noise on the patch. Then: It may be much easier to MANUALLY analyze these data than trying to decipher the results of Imatest digesting the same data…

Do you have this at hand?

Pompo Bresciani
Pompo Bresciani
Member
October 8th, 2020

So the rolling shutter at 8.7 ms is better than Sony a9-a9II stacked sensor !?!

Philip Bloom
Philip Bloom
Guest
October 8th, 2020

Yes. The A9 does skew correction in camera for stills.

Pompo Bresciani
Pompo Bresciani
Member
October 8th, 2020
Reply to  Philip Bloom

Really? I have always thought that the non existent rolling shutter was due to its super fast sensor readout speed.

William Reynish
William Reynish
Member
October 7th, 2020

I don’t care about the results with SLOG, what I want to know is how it stacks up when using Prores RAW.

Jonathan Gentry
Member
October 7th, 2020

Hi and thanks for the evaluations. You represent a unique point of common testing between cameras. All we really want to know is – how do they compare? What are the pitfalls in the response.

I want to reiterate a few thoughts:
1. We would like to see the Alexa (and possibly Sony Venice) tested as a reference using new methods.
2. We would like to see RAW outputs tested as a comparison.
3. We would like to see some method to test color accuracy in over and under.

As you can see these cameras (A7Siii, the S1H, the S1, R5, C300Markiii, C500Markii, FX9, C70) represent the vast majority of interest in Digital Cinema right now. If there were a point at which you could pull out Alexa with the above cameras and do a large comparison it would be a real hit. I think you could get plenty of support from your user-base in a test like this.

Last edited 21 days ago by Jonathan Gentry
BOUNCE
Member
October 9th, 2020
Reply to  Gunther Machu

Let’s not forget the Red Komodo 6K. We definitely need an independent test of that camera.

Last edited 19 days ago by BOUNCE
Alan Stankevitz
Alan Stankevitz
Member
October 7th, 2020

FYI, I do astrophotography / astrovideography and the A7S3 is doing something strange with their noise reduction. It eats stars. For example, shooting Slog 2 (PP7) and a sharp lens such as the Sony 20mm 1.8 stars look fine at ISO 80,000 but then bumping the ISO up to 102,400 some form of NR kicks in and the dimmer stars disappear.

This star-eating happens at various different ISO’s depending upon the picture profile. Even when picture profile is turned off it happens as well albeit at a much lower ISO.

I’m half-tempted to pick up a Ninja 5 and record externally to see if this goes away. I suspect it will since it appears to be NR that’s causing this problem.

Cliff Totten
Cliff Totten
Guest
October 7th, 2020

Sony normally has a two stage noise reduction built in. First is image processor noise reduction. It then gets a second and lighter noise reduction on the XAVC encoder chip. So,…traditionally the XAVC file is hit twice but the HDMI output only has one. RAW out on the S-III should (hopefully) have no noise reduction.

adam baker
adam baker
Guest
October 7th, 2020

You shouldn’t use slog3 for night photography. As for star eating no idea. I have ninja and shoot prores raw 4.2K and its got more noise than standard and other PP. It bypasses the NR.

Last edited 20 days ago by adam baker
Taehun Yoon
Taehun Yoon
Member
October 7th, 2020

I hope to see dr tests with external recorded prores footage. many people now know how h264/265 codecs are bad for editing and post production. and sometimes external recording shows true power of a camera.

René Rainer
René Rainer
Member
October 7th, 2020

Thx for the very informative test! Very much appreciated Gunther!

Dennis Schmitz
Dennis Schmitz
Guest
October 7th, 2020

Engineers: How much noise reduction and sharpening do you want?
Sony: YES!

Philip Bloom
Philip Bloom
Member
October 7th, 2020

Hey Gunther. Why did you do the second ISO at 16,000? The production model switch is at 12,800 now. P

Cliff Totten
Cliff Totten
Guest
October 7th, 2020
Reply to  Gunther Machu

I dont know for sure but if Sony modified the dual gain mappings and dropped the second stage (“high” 0db) to 12800 ISO, it was probably for an important reason. They might have tweaked several other sensor related things along with that too.

adam baker
adam baker
Guest
October 7th, 2020
Reply to  Gunther Machu

Can you pop on a ninja V with the 12bit log and test the 4.2k DR..

Also shooting raw is only way to get dci4k

Cliff Totten
Cliff Totten
Guest
October 7th, 2020
Reply to  Philip Bloom

Yup!…the S-III with the latest firmware lowered image second gain stage to 12800 ISO. If your camera is still at 16000 ISO, you need to update it.

Paul Corneille
Paul Corneille
Guest
October 7th, 2020

Very nice test and nothing new, as expected from a 10 bit non raw camera.
I am curious to know about color accuracy because personally I really don’t care about low light and crazy dynamic range.

Christopher Dobey
Guest
October 7th, 2020

Very curious about the positive impact of 16-bit RAW out :)

Dennis Schmitz
Dennis Schmitz
Guest
October 7th, 2020

Don’t expect anything. It’s still a 12 bit ADC in video mode. I don’t think they can enable the 14 bit mode in such a tiny passively cooked body.

Christopher Dobey
Member
October 7th, 2020

Thank you as always Dr. G! Many of us were waiting for these results.

Given that the A7S III is the first Full Frame hybrid video/photo mirrorless to output full 16-bit RAW video over HDMI, I’d be curious to see how the video quality compares to the lower 14-bit RAW still photos that it takes internally.

Lucas Mason
Lucas Mason
Guest
October 7th, 2020

It’s been revealed by several sources before that this so-called 16bit RAW is dithered from 12bit ADC readout.

Christopher Dobey
Member
October 7th, 2020
Reply to  Lucas Mason

Lucas please provide a link to these several sources. Are you referring to the ‘leaked’ Sony IMX521 CQR sensor white sheet back in late 2019? Which could do 7fps 16-bit in Readout Mode 1A but only at 7fps.comment image

Sight Sound
Sight Sound
Member
October 7th, 2020

It would be very helpful for these DR tests to be re-run utilizing the 12-bit externally recorded raw files. Theoretically these should be free of the NR.

Eugenia L
Member
October 7th, 2020

Thank you for these tests. Please consider renting an Alexa LF, because we have no frame of reference for these DR tests. You see, everyone said in the past that Alexa has 14.6 stops of DR, but who knows what kind of test they used. I’d like to see what your test says about that, so we can then better judge the consumer cameras too.

Regarding this test, I wish you had tried the 10bit 4:2:0 h.265 version at 100mbps too. You see, the all-intra h.264 version is not necessarily the best, your test hit the macroblocking that is usual wit h.264. As for why I’m suggesting 4:2:0 and not 4:2:2 for the above, is because Sony doesn’t give you more bitrate for 4:2:2, while it does require 20% more than 4:2:0 to store the color data. They give you more in 25p and 30p, but not in 24p (cripple hammer from Sony here). So at 100mbps h.265, the 10bit 4:2:0 is the best quality you’re going to get in that camera. Not to mention that GPUs don’t support 4:2:2 at all (only the iOS chipsets do, and a new Intel laptop gpu, all the other ones don’t — decoding falls to CPU). So I wish that combo was tested too.

Last edited 21 days ago by Eugenia L
Valentin
Valentin
Guest
October 7th, 2020
Reply to  Eugenia L

But the the testing was done in 25p with 250Mbit/s. When I select h.265 Sony doesnt let you record 25p… only 50p or 100p available. Is there any reason for not providing XAVC-HS h.265 in 25p?

Philip Bloom
Philip Bloom
Member
October 7th, 2020
Reply to  Valentin

Shoot XAVC S I. Don’t bother with HS

Silton Buendia
Silton Buendia
Member
October 6th, 2020

Take this test as a grain of salt. Not all test are created equal as not all cameras are either. Some do better in the high range while others do better in the low. Also you can’t expose them the same if measuring DR in log because as I stated cameras are calibrated different. Sony for example is set to handle highlights better while another camera that might score higher in this text won’t retain highlights as good. Also because of this Sony’s benefit from a little bit over over exposure where as other camera need to be exposed less. Also from side by side comparison In numerous test it’s quite clear that this test might be severely outdated because visual DR perception has appeared to be better in both the Sony FX9 and A7s3 then Canons

Last edited 21 days ago by Silton Buendia
Paul Bennett
Paul Bennett
Guest
October 6th, 2020

Dynamic range <= Panasonic S1H, latitude < S1H, based on CineD tests of both, with S1 and S5 very close to S1H.

S1 rolling shutter is better, 7.4, in FHD APS-C mode. Not sure about the S1H as CineD didn’t test that mode.

I think the Panasonic image quality is more filmic, and seems to be much easier to grade as almost every A7S III video I’ve seen is too contrasty and has the shadows crushed.

So A7S III looks to me to be pretty much about the continuous autofocus. If you don't need that I think Panasonic is the winner.

Will Panasonic ever do phase detect or dual pixel autofocus? So odd that they have not.

Last edited 21 days ago by Paul Bennett
Paul Bennett
Paul Bennett
Guest
October 6th, 2020
Reply to  Paul Bennett

Forgot to mention Panasonic wins on stills as well with 24mp sensor and 96mp high res mode.

Lieutenant-Dingo
Lieutenant-Dingo
Guest
October 6th, 2020
Reply to  Paul Bennett

doesn’t the S1H have quite aggressive noise reduction? I believe this contributes to its slightly higher dynamic range. As for image quality, I think that has more to do with the way others on the internet have exposed and graded footage, opting to not overexpose the footage as was ideal on previous models. The raw dynamic range of all these hybrid cine cameras must be similar if they all have 12bit ADCs, though I have yet to find information regarding the A7siii’s ADC bit depth in movie mode.

Paul Bennett
Paul Bennett
Guest
October 6th, 2020

In 6K or 5.9K the S1H has 12.7 stops and yes Gunther thought that was probably due to noise reduction. But in 4K the S1H has 12.3 without the aggressive noise reduction, pretty much the same as the SIII’s 12.4.

Which brings up another point – Panasonic has 6K :-)

I’m trying to decide how much Sony’s excellent autofocus matters, vs Panasonic’s decent but not so excellent autofocus.

Dennis Schmitz
Dennis Schmitz
Member
October 7th, 2020
Reply to  Paul Bennett

12.7 stops was with the strong temporal filtering similar to the A7s3 which literally rendered the image unusable.

With firmware 2.0 and noise reduction set to -1 the image is basically flawless.

Dennis Schmitz
Dennis Schmitz
Member
October 7th, 2020

The S1H had horrible noise reduction similar to the A7s3 until firmware 2.0 and the Noise Reduction -1 setting.

Now the image looks super organic with loads of texture and is free of motion artifacts.

And all these cameras have a 12 bit ADC video mode indeed.

Eugenia L
Member
October 7th, 2020
Reply to  Paul Bennett

The green skintones people get with Sony is due to Sgamut/cine. The Sgamut is massive, and it does not shrink properly to rec709 on any video editor (not even in Resolve). As for the Luts Sony offer are terrible and haven’t been updated in years. So colors don’t match up well. If you want a better color out of Sony, you need to change your gamut to rec709 on your slog2/3. Or just use HLG3 with rec709 (not bt2020 as most recommend — green will still be there).

Particularly in 8bit, it’s color suicide to use slog2/3 (slog3 at that mode is like working with 6 bit color).

Personally, in the past I only appreciated the Alexa and Canon colors, and didn’t like at all Sony colors. But after doing tests with the rec709, and setting the right exposure and right WB each time (very important for sonys), the image was very gradeable and very easily made to look like actual film. None of this weird yellowness on the leaves anymore, or the greenish skintones.

Mike
Mike
Guest
October 7th, 2020
Reply to  Eugenia L

How do.you know so much? You a colorist?

Also wouldn’t the 600Mbps at 422 work? I do believe a7siii allows that at 24fps… on a7s3 fyi

The DP Journey
Guest
October 6th, 2020

Thanks guys! Very helpful as always.

Wolfgang Andrä
Wolfgang Andrä
Guest
October 6th, 2020

Thanks for your test! I think you should also test the dynamic at 12.800 ISO. Because Philib Bloom pointed out, that his production units second base ISO start now at 12.800. Exactly after 4 Stops over base ISO. The same rule is true for each profile. Maybe that’s the same in your unit.

Valentin
Valentin
Guest
October 6th, 2020

According to Philip Bloom and my own testing, the second native iso curcuit starts at 12800 with the productio unit, not at 16000. You can clearly see it when switching from 10000 to 12800, even on the cameras lcd. Could you redo the dynamic range test with the correct 2nd iso step?

Valentin
Valentin
Guest
October 6th, 2020
Reply to  Valentin

Also how did Gerald Undone reach 13 Stops with his test? He is using the same device to read the DR

Dennis Schmitz
Dennis Schmitz
Guest
October 6th, 2020
Reply to  Valentin

Different version of IMATEST using slope-based instead of quality-based (SNR) measurements.
https://www.imatest.com/solutions/dynamic-range/

Slope-based leads to even more unrealistic results.

Lieutenant-Dingo
Lieutenant-Dingo
Member
October 6th, 2020
Reply to  Dennis Schmitz

Am I right in saying that Gerald uses the newer 2020.1 version (which is quality based) while CineD uses the older version 4.2?

Dennis Schmitz
Dennis Schmitz
Guest
October 6th, 2020

I think the older version is quality-based while the newer one is slope-based.

About slope-based: “…but it should be interpreted with care because SNR can be extremely low (often well under 1 = 0 dB) in the the dark portion of the range. This can result in measurement inconsistencies as well as overly optimistic measurements, since no meaningful image is visible (it’s all noise) when SNR < 1. Slope-based DR can be falsely increased by tone-mapping and by medium-range flare light (bleeding from light chart patches, visible in the Image Statistics images below). For this reason WE DO NOT RECOMMEND IT."

Lieutenant-Dingo
Lieutenant-Dingo
Guest
October 6th, 2020
Reply to  Dennis Schmitz

Imatest also states “Flare light was not an issue with the high-quality DSLR/mirrrorless lenses we tested in the past, but it has become a major factor limiting the performance of recent low-cost lenses intended for the automotive or security industries. We have seen examples of how flare light can improve traditional DR measurements while degrading actual camera DR.

Our approach to resolving this issue is to limit quality-based DR measurements (the range of densities where SNR ≥ 20dB for high quality through SNR ≥ 0dB for low quality) to the slope-based DR. This works because, for patches beyond the slope-based limit (where the slope of log pixel level vs. log exposure drops below 0.075 of the maximum slope):

Contrast is too low for image features to be clearly visible.
Signal is dominated by flare light, which washes out real signals from the test chart; i.e., the “signal” is an artifact, not the real deal.
Limiting quality-based DR in this way significantly improves measurement accuracy, and perhaps more importantly, can help prevent inferior, low-quality lenses being accepted for applications critical to automotive safety or security,”

which seems to contradict the statement they made above as to the reasoning for moving to slope based DR testing and its effects on reliability and accuracy.

Gergö Nyirö
Gergö Nyirö
Guest
October 7th, 2020
Reply to  Valentin

yes please!

Greg Greenhaw
Greg Greenhaw
Guest
October 6th, 2020

Can you do some with pro res raw with noise reduction and slight highlight adjustment to me more consistent with a real world use case?

Dennis Schmitz
Dennis Schmitz
Member
October 6th, 2020
Reply to  Greg Greenhaw

I’m expecting 11.5 stops for the S5/S1H and A7sIII in ProRes RAW which is like the possible limit with 12 bit ADC and without heavy noise reduction.

S1H/S5 may have a slight advantage after downscaling to 4K.

1
1
- 3
Filter:
all
Sort by:
latest

Take part in the CineD community experience