Review: Sony NEX-FS100 “Super35” LSS AVCHD Camcorder

Josh Dahlberg: | February, 29, 2012

Thank you Adam - a comprehensive, balanced and insightful review as always.  I’ve been in the industry ten years now and right from the start have lapped up your camera profiles - they’re unmatched anywhere on the web.  The detail and range of info (from handling to chart interpretations) is just brilliant.  This one is particularly useful for me in determining whether the FS100 is “good enough” or if I should hold out for a C300.  Much appreciated.

Josh

ghosteater: | February, 29, 2012

Awesome! Very informative and appreciated. Thanks

tunny: | March, 11, 2012

A very well written and thorough review. I do have to comment on one thing though. You say: “Its (the AF100s) sensor is about 75% the size of the FS100’s, so it has roughly comparable depth of field ...”

That may be true of the total sensor areas, but it’s neglecting that the four-thirds sensor is 4:3. Used in the AF100 it gets windowed to 16:9, which reduces the effective area by 25%, so that when you compare ACTIVE areas between the two cameras the AF100 is approximately 50% that of the FS100.

For depth of field purposes, the FS100 can be considered to have a full stop advantage over the AF100.

Don Downie: | March, 11, 2012

Thanks for the great article, Adam.

One quick question:

You wrote that “The FS100’s timecode generator can be jam-synced over LANC for multicamera shoots.”

I own an FS100, and am hoping to try syncing with another one on a shoot soon.  I can’t find any documentation of this feature, though.  Do you mean that I can just plug a 2.5mm male to male stereo cable between the LANC ports on two FS100s, and one of them will slave timecode?  Or do I need to do this via a LANC controller of some kind that feeds both cameras?

Thanks again,

DonDownie

tunny: | March, 11, 2012

I agree - a very good article.

One point of clarification is regarding the relative sensor sizes between the AF100 and the FS100.

It may be true that for OVERALL area the FS100 is only 75% than the AF100, but because the ffour-thirds sensor is 4:3, 25% of that is lost in the AF100 in 16:9 mode.

Hence I believe the FS100 is about twice the size of the AF100 when you consider ACTIVE area - a full stop difference.

tunny: | March, 11, 2012

Correction, and my apologies - the above should have read that for overall area, the AF100 is 75% that of the FS100. But for active area, it’s only about 50% of the FS100, due to the AF100 only using a 16:9 window of the 4:3 chip.

Adam Wilt: | March, 13, 2012

Regarding sensor size and depth of field: the FS100’s active area is 23.6x13.3mm, the AF100’s is 17.8x10.0mm. The AF100’s sensor is indeed 75% the linear size of the FS100’s, and 56% of the area.  Depth of field scales with linear dimension, not area; there is about a one-stop difference between ‘em: for the same angle of view, you’ll need around f/2.8 on the AF100 to match the depth of f/4 on the FS100.

I consider this to be “roughly comparable” when compared to the difference between the FS100 and a 2/3” or smaller sensor, where you’ll need three stops wider or more on the small-sensor cam to match the depth of f/4 on the FS100. Sorry I didn’t spell that out.

As to area scaling, yes, the AF100’s 56% active area compared to the FS100 translates, all else being equal, to a one-stop sensitivity advantage for the FS100.

All else is far from equal, though; the AF100 appears to line-skip (and pixel-skip within a line) to generate an HD image from its 12 Mpixel sensor, so it (again apparently) loses more active area to that skipping. Further, Panasonic’s sensor aren’t directly comparable to Sony’s sensors in terms of noise and sensitivity even when they have equal photosite counts and active areas. Most critically, the signal processing differs markedly between the two cameras: the AF100 appears to be using substantial 3D noise reduction, such that its apparent noise level doesn’t change at all with gain boost—only its character changes.

Increasingly as we move into the era of cheap-as-chips (grin) DSPs, what comes out of the camera bears less and less resemblance to the “pure and unprocessed” signal emanating from the sensor… at which point it’s worth comparing cameras more with reference to the images they make and less with reference to the fundamentals of geometry and basic physics. It’s getting to the point where I’m about ready to answer any question about comparing sensors from different manufacturers in different cameras with “mu”, the answer to the koan, “does a dog have Buddha-nature?”!

tunny: | March, 14, 2012

Thanks for the clarification - I had understood you meant area when referring to 75%. I think we’re both in agreement that the FS100 sensor is about 100% larger than the AF100 in AREA terms, and that makes about a stop difference for relative depth of field comparisons.

I do think that it’s quite a significant difference though. Yes, they both offer a significant difference in depth of field compared to a 2/3” sensor, but it’s 2 stops for the AF100, 3 stops for the FS100.

Yes, if all else was equal it should also mean a stop more sensitivity for the FS100 - but as you say, the FS100 is far better than just one stop in practice. The only explanation I’ve heard that fully explains the observations is that the AF100 is fundamentally reading the sensor in a similar way to the C300 (direct read of R,G,B from 2x2 Bayer blocks) *BUT* with the big difference that it only reads every other block horizontally and vertically, so only one block in four, or only four photosites out of every sisteen.

That would be expected to give a final output resolution of a quarter the native chip dimensions, symmetry for horizontal and vertical characteristics, and quite high luminance aliasing but little or no chroma aliasing. All exactly as measured. (For a 16:9 12 megapixel sensor, I’d expect dimensions of around 4800x2700, so a predicted output resolution before aliasing of a quarter that - 1200x675 (675lpph). That correlates well with the RB-SRW chart you show for the AF100, especially looking at the res trumpet with mainly horizontal lines.)

When the AF100 was developed it’s competition was against DSLRs, and it’s readout system is undoubtedly better than any DSLR of the time for video, even allowing for only one in four photosite being actively used.  The real lesson seems to be there is a big gulf between purpose designing a chip for video, and “making do” with what is effectively a still camera sensor.

Don Downie: | March, 14, 2012

Adam, any clarification on how to jam sync over LANC?  I would love to know the method for this, if it is in fact possible.  Thanks!

Adam Wilt: | March, 14, 2012

tunny: I won’t argue whether a 1-stop DoF difference is significant. My only point was that a 3-stop difference is about 3x as significant as a 1-stop difference!

Interesting theory on the AF100 readout. It does seem to predict the observed results, on that camera as well as on the GH1 and GH2. Have you any pointers to that explanation?

“The real lesson seems to be there is a big gulf between purpose designing a chip for video, and ‘making do’ with what is effectively a still camera sensor.” +1!

Don Downie: A Sony chap told me, “It is possible to synchronize Rec START/PAUSE and Time Code Preset via the LANC remote port”, which is what I based my statement on.

I’m trying to get confirmation of this, but so far all I’ve got back through official channels is. “You can definitely put the cameras side by side and using the IR controller reset all to zero and start the TC generator. TC generator must be previously set to free run.” This is the common-start technique long available with any IR-controlled camera, and one I’ve employed profitably on multicam shoots in the past. It’s typically good for getting TC synced to within a couple of frames.

If I hear anything more definitive soon, I’ll update these comments; in the absence of confirmation I’ll edit the article to remove that statement. Stay tuned…

Oh, BTW, Sony has announced that the firmware upgrade for the FS100 should be available on 22 March. (A tip of the hat to sonyalpharumors.com for the news!)

Don Downie: | March, 14, 2012

Thanks, Adam.  Hope I didn’t come off as cheeky—I’d be really excited to have a hard-wired way to sync these. 

If I get a chance, I’ll try to test syncing via LANC port and a simple cable.

I’ve had mixed results in the past using IR remote to sync cameras, but that was on Panasonic HVX200s, which had some timecode peculiarities (in 24pN).

Please keep us posted if you hear more from Sony.  I’ve got my breath held for that firmware upgrade.

tunny: | March, 14, 2012

“Interesting theory on the AF100 readout. It does seem to predict the observed results, on that camera as well as on the GH1 and GH2. Have you any pointers to that explanation?”

It was something I read quite a while ago, but I think was speculative rather than from an official source. I seem to remember it giving other pointers as evidence as well as sensitivity, aliasing and resolution - power consumption and simplicity of implementation amongst them - and it explains why the resolution charts are symmetrical. (Unlike DSLRs with simple line skipping.)

At the time, the idea of directly reading out 2x2 blocks to derive video wasn’t generally known about (AFAIK). It was only when the C300 came out with a description of the same fundamental technique (albeit with a sensor which could use the technique to it’s full with no skipping, and give full 1080 resolution with no up or down conversion!) that I remembered what had previously been suggested as an explanation for the AF100. The C300 proved the 2x2 block readout theory was valid.

Adam Wilt: | March, 14, 2012

Actually, 2x2 direct readout has been around at least since 2000 when I wrote about color in CCD cameras, at least in interlaced dual-row readout of CMY+G sensors (see the spec sheets at http://www.sony.net/Products/SC-HP/pro/image_senser/color_video.html for current-product implementation; the spec sheet URL I originally referenced having long since gone missing).

In ye olde days, the dual-row readout trick worked well without requiring twice the output resolution in the sensor, though if the sensor has less than 2x the horizontal resolution, H res would of course suffer (for analog color-under camera or security cameras, that wasn’t a big deal anyway!). For progressive scan with full resolution, one does need a full 2x in both H and V, like the C300 uses.

tunny: | March, 19, 2012

Adam - I’ve been having a closer look at the charts you shot and what most interests me is the diagonal on the FS100. The zone plate shows strong alias circles (with a null at around 960lpph) but they seem to be prominent on the DIAGONALS. This is what I would expect with the photosite rows at 45deg to to horizontal/vertical. It looks very like a normal zone plate rotated through 45deg.

For the last year, I’ve believed the F3 and the FS100 to share the same sensor - but I’m now having my doubts. I’m sure the F3 sensor is conventional Bayer and about 3.3 megapixel, but I now suspect the FS100 may be different. This may explain why Sony initially gave details for the F3 but now do not soecify any details for the F3 or FS100.

It’s a big leap - but I wonder if we may be seeing on the FS100 the Q67 sensor that was talked about a year or more ago….....? The low power consumption of the FS100 seems to point to a more simplified readout system than the F3, and in that case the lack of coloured aliasing would indicate a much higher photosite count than 3.36 megapixel.

Adam Wilt: | March, 19, 2012

The Sony folks keep insisting it’s the same sensor in both the FS100 and the F3.  The specifications simply state:

Image device
“Exmor” Super35 CMOS sensor 23.6 mm

tunny: | March, 19, 2012

I confess I’m puzzled - I’ve just looked up the manual online and you’re quite right regarding what Sony state there.

As example to show why I’m confused, I’m theorizing a situation (and assuming the 2456x1372 Bayer) where it is illuminated by an ideal pattern of 1372 horizontal lines, half white, half black. Let’s further assume the white lines fall on the Bayer rows RGRGRG etc. That means that NO B photosites will receive illumination, so no matter how the chip is READ, there can’t be ANY blue in the output? Shift the pattern down a line and the illuminated rows will all be BGBGBGB etc, so in this case no chance of any red in the output.

Hence I’d expect the coloured aliasing to be inherent to the sensor geometry rather than readout method? I can’t see any mechanism that could avoid coloured aliasing around the 1372 lpph figure on a zone plate.

I looked at the NX5 zone plate you mention and couldn’t help think that the general form of it did look similar to the FS100 in terms of the predominant alias circles being on the diagonal - not on the axes. (Though in the case of the NX5 the null is closer in to the centre than for the FS100.)

Adam Wilt: | March, 19, 2012

Consider also the use of optical low-pass filtering and how it works to spread light rays across multiple photosites. Without an OLPF it is theoretically possible to generate the sort of pathology you describe, but in practice it’s exceedingly rare to actually get that perfectly awful situation. With an OLPF, it’s even rarer.

Here’s an interesting demosaicing comparison site: http://thedailynathan.com/demosaic/ Have a shufty ‘round the web for other such demos and comparisons, too. Different demosaicing processes, given real-world images, can do surprisingly different things with fine details.

tunny: | March, 20, 2012

Regarding the pathology I described, then it is showing up extremely well on the F3 chart - “blotchy” patches of yellow and cyan on vertical and horizontal lines on the chart from around 1150-1600 lpph (centred on the 1372). That’s exactly what I’d expect, it may be rare in “real life”, but there will be points on such a chart where it would be expected.

I confess I hadn’t really previously considered the OLPF. But looking at the FS100 chart again with that in mind, then I’m struck by the aliasing in the 45deg diagonal segment. The converging response is as expected up to about 1000 lpph, but then the lines start to diverge (aliasing) and this pattern is clear right through to beyond 1750 lpph, when a hint of colouring starts to show. If the OLPF was having much effect in suppressing the pathology I describe around 1372 lpph, then why such a level of luminance aliasing?

I’m inclined to conclude the OLPF is not having too much effect at this spatial frequency. And so we come back to the previous point - if no B photosites are giving an output, how can any algorithm reconstruct to a monochrome image? (And one which represents luminance only aliasing at that!? )

mrak1979: | April, 03, 2012

Great review, but I’ve heard this camera suffers from really low highlight protection and that highlights blowout way too quickly. What are people’s thoughts?

thx

———————————

Adam Wilt here, jumping in via the comment-editing interface, since our CMS is refusing to let me post any comments through the normal channels (!) (sigh):

Its highlight protection, a.k.a. its knee function, is in my experience no better or worse than any other sub-$20K Rec.709 camcorder or camera with a linear knee.

The S-curved knee on the new Panasonics, like the HPX250, AC160, and AC130 does a slightly more natural job of rolling off the highlights than a linear knee does, but for a substantial improvement (IMHO) in highlight quality, you’ll need something with a “cinegamma” or “hypergamma”, like the EX3, EX1R, or other XDCAM EX cameras have.

Fortunately, the just-announced NEX-FS700 has cinegammas—the first time they’ve migrated down-market to the NXCAM line.

If that won’t do it for you, well, you can always got get an Arri Alexa… 

———————————

tunny: | April, 11, 2012

I hadn’t really expected to add anything to my last comments - but in the last three weeks the FS700 has arrived, together with a sensor with far higher pixel count than 3.3megapixels. In light of my previous thoughts about the FS100, I would be very interested to see a chart from it shot in a comparable way to those you have done for the FS100.

Could there just be a chance that the FS100 and 700 will have certain very similar “fingerprints”.....?

It would explain a lot…..

Mike N: | April, 24, 2012

Hello Adam,

Found your review of the camera a couple of days ago which is very helpful and extensive, thank you. Hope you still read this and can answer me some questions.

You said that the Cine gammas on this camera depress the midtones a bit. So if for example in the 709 curve the input signal is 40 and it comes out 60 on screen then for the Cinematone the 40 goes down to maybe 50 ? Am I getting this right ? Should I expose faces as I would with 709 gamma or compensate for the lower midtone and crank the exposure up or add more light to bring the face back up, or should I do this through grading.

Secondly, compression in the highlights results in more information being added in the image as the cost of awkardly looking colors and posterization if we exagerate the knee. What about the shadows ? To get more detail from them shouldn’t I expand them (dialing in a posivive level +) rather than compress them ?

thanks,
Mike

Adam Wilt: | April, 25, 2012

Mike N: I’m not sure why an input signal at 40 would come out at 60, but in general, for the same exposure setting, the Cinematone gammas will have the same scene midtone value (say, a gray card) reproduce at a lower level onscreen than it would in 709 gamma.

Where you set exposure for faces, and which gamma you choose, isn’t something allowing a simple one-line answer. The different gammas render scenes differently, and you use them as a tool to affect how you choose to have the scene rendered, along with color balance, knee, etc.

Knee compression doesn’t really “add more” to the image, but it squeezes more highlight values into the same tonal scale by reducing the separation between those highlight tones. while it captures more scene brightness into the highlights, and may include details that would otherwise be blown out, it actually reduces the discrimination between details (as defined by variations in brightness) in those highlights. Put another way, you may capture more brightness levels in the highlights, but you reduce contrast between them in doing so.

As to shadow detail, the black gamma setting isn’t really the inverse of the knee; it can’t “capture more black” the way the knee can “capture more white”. Rather, it varies the contrast in the existing shadow detail (at the expense of some midtone contrast; TANSTAAFL!). Yes, if I remember correctly, you’d increase the black gamma number to expand the shadow detail and make more things visible therein that might otherwise all be smushed together, tonally.

Mike N: | April, 25, 2012

Thank you Adam.
As for the input output signal it was my understanding that this is what gamma curves do actually. They are not an actual straight line in the same way our eyes see 18% gray as a halfway between black and white, hence a 50 output for a 18 input. I think these gammas work the same especially the Cinegammas and more so the Log curves found in the high end cameras, they have a steep curve in the blacks and midtones and after that they are very shallow at the end for the highlight. I might be wrong obviously but this is what i’ve come to understand from all the graphs and charts of gammas.
thanks.

Adam Wilt: | April, 25, 2012

Mike N: Ah, now I see what you’re saying. Yes, an 18% gray card normally comes out, in a Rec.709 gamma, at around 42% brightness (the display gamma is normally a rough inverse of the capture gamma, so that 42% brightness value gets displayed at around 18%, which is why the whole reproduction chain works!).

If you look at the grabs on page 2, you’ll see how the tonal values change with different gamma settings. The Cinematones depress midtones about 10 units compared to normal gammas, with Cinematone 1 S-curving shadows a bit, and Cinematone 2 S-curving both shadows and highlights.

Mike N: | April, 25, 2012

Thanks for taking the time and explaining this. I’ve had these questions in my head for quite some time now and they badly needed clarification.
So…to put this to rest:
The camera, the sensor actually captures data linearly, so for a 18 gray is sees a true 18 gray. After this a gamma is applied which raises this to about 42% brightness. Finally we view this on the camera display or on a monitor which has an inverse gamma (1/2.2 or something like that for “normal” LCDs right ? ) which brings back the 42% to 18% BUT because our eyes are nonlinear we actually see that as about 50 % bright ?
This of course for a 709 gamma which is very closely designed for tv and broadcast, the Cinegammas are darker, like you said with about 10 units, so 42% becomes something like 32%.

Finally on this camera the Cinematone 2 seems to capture the most details out of the shadows and highlights although i’ve seen that applying a more aggressive knee to he 709 gamma (about 80 with the slope set to minus) brings back lots of highlight detail.
thanks again.

Adam Wilt: | April, 26, 2012

For the most part, that’s how it works. More at http://poynton.com/notes/colour_and_gamma/GammaFAQ.html and i,f that doesn’t slake your thirst, Poynton’s book is worth reading through.

Mike N: | April, 26, 2012

This is great. Thank you. I’ll surely read it through.