Canon 5D Mark III - 3.5K and 4K raw video with Magic Lantern

[Shirozina]

7 minutes ago, Andrew Reid said:

.

And don't get me started on 8bit "uncompressed" HDMI... It is compressed as hell! No point in recording it. Especially not as 10bit ProRes.

On all cameras ? - Not my experience on the A7s and A7r2 where it's much cleaner than the internal codec with regards to being able to work with Slog-2 and not get banding in blue skies even with quite aggressive contrast and exposure changes. 

[maxotics]

14 minutes ago, Andrew Reid said:

10bit has more advantages than just less banding my friend

Then please prove it to me, or show someone who has.  I tried and couldn't do it (but recognize I may have screwed up!).   Show me a frame that was shot in 8-bit video and one in 10-bit video, and then a best effort to recover detail at the high or low end of the image.  

16 minutes ago, Andrew Reid said:

In rec.709 your dynamic range is limited in 8bit because you simply run out of room

It isn't limited because in rec 709 you have enough color space to show every color a human eye is able to discern without changing pupil size.  It is limited in that it can't capture the full DR of a sensor.

 

18 minutes ago, Andrew Reid said:

That's because with 0-255 you have regions of large variation in brightness crammed into the last 5 or 10 shades, so things like a sky looks like crap, no accurate colours in it

Only if you're trying to "re-expose"/grade an image that isn't already perfect within that 8-bit space.  The root problem is most camera record high dynamic range images into display-ready data spaces.  I mean, I agree with you.  Which is why you really have to nail exposure and lighting long before you aim your camera at anything

21 minutes ago, Andrew Reid said:

LOG is a way around this to a certain extent but is also a form of compression

Sorry, compression always bothers me because it's a data/computer term which implies that if done correct you don't lose information.  A compressed word document is as good as an uncompressed one.  In visual data, LOG is more of a data distortion, to me at least.

23 minutes ago, Andrew Reid said:

So in short, 10bit = more dynamic range.

In RAW yes, compressed video, no.  Again, I would LOVE to see some real-world proofs.  Not pixel peeps of banding on some wall.  Two videos played, one in 8-bit and one in 10-bit where you can see a worthwhile improvement in DR.

25 minutes ago, Andrew Reid said:

Also 10bit raw = more dynamic range, plus more control over colour and white balance as it isn't baked in

Yes, because you're reading data from a sensor that isn't trying to fit into an 8-bit space for recording.  

28 minutes ago, Andrew Reid said:

I'd say the image quality at 10bit is very close to 14bit... 12bit even closer as to makes barely any difference. 14bit has a bit more accuracy in the highlights. Shadows still look AMAZING in 10bit, like the Blackmagic raw cinema cameras but less noisy.

Yes, less noisy because you have a larger sensor, anyway, with shallower DOF looks pretty freakin' nice.  This is the point I don't know why you don't hammer in on.  10-bit RAW IS amazing!!!!

[Andrew Reid]

Ok I take on the challenge.

I'll shoot 10bit GH5 and compare it to 8bit GH4

And I'll shoot some more 10bit RAW on the 5D Mark III and compare it to 8bit 4K on the Sony cameras.

I thought Blackmagic already showed that 10bit was superior... Problem was maybe their sensors were a bit behind, so shadows were a bit noisier and so on, compared to an A7S or the like.

[Shirozina]

Wouldn't just switching between 10 and 8 bit on the same camera be a better test - easy to do on the BMPCC at least.

[Andrew Reid]

1 hour ago, maxotics said:

Only if you're trying to "re-expose"/grade an image that isn't already perfect within that 8-bit space.

Perfect isn't an option all the time... Nature doesn't always give you perfect lighting.

If a controlled interior scene is lit for the narrower dynamic range of a Rec.709 8bit image you can get it looking good, as the guys with the GH2 showed in the famous Zacuto shootout that had Francis Ford Coppola fooled.

If you try to pack too much dynamic range into the frame then colour is going to suffer with 8bit.

That is why the 1D C has good colour but not 14 stops DR

And it is why 14 stops DR in S-LOG doesn't have good colour

And it is also why Canon LOG on the 1D C never quite regains the punch of the standard picture profiles if you want to keep your extra highlights and shadows intact.

It is all the fault of 8bit.

I foresee BIG gains for Sony colour if they do the entire imaging pipeline 14bit -> 10bit on a future A7S3, along with the improvements to the colour science evident from the A9's JPEGs.

Quote

 The root problem is most camera record high dynamic range images into display-ready data spaces.  I mean, I agree with you.  Which is why you really have to nail exposure and lighting long before you aim your camera at anything

Isn't that what I am saying?

Quote

Sorry, compression always bothers me because it's a data/computer term which implies that if done correct you don't lose information.  A compressed word document is as good as an uncompressed one.  In visual data, LOG is more of a data distortion, to me at least.

Well there's lossless compression which means you don't lose any visual information, the structure of the file changes, but the end result is the same.

Compression can be lossy... meaning you throw stuff away in the image.

LOG is lossy in 8bit, you are throwing away fine gradation and colour information to make room for more dynamic range.

Quote

In RAW yes, compressed video, no.  Again, I would LOVE to see some real-world proofs.  Not pixel peeps of banding on some wall.  Two videos played, one in 8-bit and one in 10-bit where you can see a worthwhile improvement in DR.

Honestly it's easy to see

When you tweak the shadows and highlights in 10bit RAW on the 5D Mark III it maintains the natural, realistic still-life look of the shot, smooth roll off between two subtle tones in the shadows like a dusk sky and the street... whereas in 8bit on a lossy camera it turns to Digital Sick™

 

[maxotics]

16 minutes ago, Andrew Reid said:

Ok I take on the challenge.

I'll shoot 10bit GH5 and compare it to 8bit GH4

And I'll shoot some more 10bit RAW on the 5D Mark III and compare it to 8bit 4K on the Sony cameras.

I thought Blackmagic already showed that 10bit was superior... Problem was maybe their sensors were a bit behind, so shadows were a bit noisier and so on, compared to an A7S or the like.

AWESOME!  I tried to do comparisons using the Sony X-70 but arrived at the conclusions above.  I also want to clarify that I'm not saying 10-bit doesn't have an advantage.  Obviously, any improvement of the image is welcome.  It was just disappointing to me because I wanted to capture more DR.  It took a me a while to understand why there wasn't much of an improvement in 10-bit.  I didn't make any sense to me.  

BTW, I'd buy your guide if you had one for my camera   Why can't you just sell a master RAW for Canon guide and include most cameras.  When friends ask me how to do it and I'm not in that mode I have to send them to the ML forum (good luck to them  ).  As I've mentioned elsewhere, the 7D is a monster set-up.  I got one 6 months ago for $400 and a 17-50 2/8 for $250.  But I don't know what the best/stable ML build is.  I may be behind the times.  If you had a guide for all the cameras I'm pretty sure I would buy it every year.   Brush up the 50D guide

It's a sh_t load of stuff to keep up with but only you have enough contacts/expertise to make it possible.  

 

[Andrew Reid]

What 10bit from what camera is not giving you the expected dynamic range improvement? What have you been testing? Name names!

[Shirozina]

10 bit won't give you a dynamic range improvement - 8 or 10 bit is the output file bit depth. A camera can have a 14+ stop DR and output this into an 8 bit jpeg image that will look very good - anybody who is working with in stills with RAW capture can tell you this esp if they have a good MF digital back which are even 16 bit capture and you can create absolutely stunning 8bit output files full of subtle graduations in tone and colour that you can only dream of with a DSLR. 8 bit is absolutely fine as an output file .................as long as you nail everything in camera and your cameras internal processing engine isn't compressing the hell out of the image and throwing away image data which unfortunately with internal codecs which are designed for broadcast and not storage it certainly is. Massive lossy data compression is the enemy we are up against - 8 vs 10 bit output is a red herring IMO. The ML RAW hack simply removes this obstacle and the benefits are clear to see. 

[Andrew Reid]

But 8bit IS itself lossy data compression.

Look at the difference between 8bit raw and 14bit raw in terms of dynamic range.

10bit is very important as an acquisition format

If your capture format is 8bit from the sensor to the card then you are losing dynamic range.

The only reason an 8bit JPEG works is because all the image processing is done in 14bit and the sensor is 14bit

As a delivery format on our screens, 8bit works ok (although there is a reason HDR displays need 10bit)... but the capture can benefit from 10bit and higher... for more dynamic range, less compression and better colour.

The reason so many people can't tell the difference between 8bit and 10bit is they are both so heavily compressed or hacked down on cameras like the FS5.

[Shirozina]

Where can you get 8 bit RAW these days? - AFAIK all cameras we are using today capture at 14 bit at the A/D after the sensor. 

[maxotics]

6 minutes ago, Shirozina said:

Where can you get 8 bit RAW these days? - AFAIK all cameras we are using today capture at 14 bit at the A/D after the sensor. 

Nowhere that I know of.  10- and 12-bit RAW, far as I know, is just that 14-bit value truncated.  It works, because as we all know, those values are generally thrown out anyway.

[Andrew Reid]

24 minutes ago, Shirozina said:

Where can you get 8 bit RAW these days? - AFAIK all cameras we are using today capture at 14 bit at the A/D after the sensor. 

With Magic Lantern

They chop the bits off at certain ISOs, giving 8bit RAW

Go and test that and compare to the 14bit raw

Then you will see a direct example of the advantage of higher bit-depths like 8bit vs 10bit.

14bit capture from the sensor is on nearly all cameras (some are 12bit)... I am talking about the recording format and image processing pipeline.

[Shirozina]

7 minutes ago, maxotics said:

Nowhere that I know of.  10- and 12-bit RAW, far as I know, is just that 14-bit value truncated.  It works, because as we all know, those values are generally thrown out anyway.

Sony A7xx cameras can get way more DR in video  than it's theoretically possible to do if the data is being truncated to 8bit after the A/D stage and before it gets compressed. 

[Andrew Reid]

Yes but that DR is at the expense of something else.... Which is what I am telling you guys all along!

[maxotics]

21 minutes ago, Shirozina said:

Sony A7xx cameras can get way more DR in video  than it's theoretically possible to do if the data is being truncated to 8bit after the A/D stage and before it gets compressed. 

That's not how I understand it.  Again, I'm a hard-data guy.  8-bit is 8-bit is 8-bit   Larger sensor cameras can capture more theoretical DR, which is born out by DxO and Bill Claff tests.  There is more DR from a A7S, especially in low light, because its pixel to area ratio is much higher than MFT sensors with more pixels than are necessary for video (but important for photography).   

One of traps is assuming an 8-bit value from compressed video matches an 8-bit value from RAW.  There are NO real 8-bit color values in compressed videos.  They are pseudo-generated from a 24-bit color value which itself is saved in a 6 color matrix in some 4:X:X form.   In RAW, the 14bit value is a REAL value of a red, green or blue filtered pixel/sensel.  There are no real colors in RAW, or dynamic range.  They are just physical readings of light hitting silicon.   In sense, you HAVE to pick a black level in all 8-bit compressed video.  When shooting 10-bit RAW, you don't need to do that so you can be off a stop in your exposure and set your black level to make a nice image, whereas in 8-bit compressed video, you're more locked in.

So even though you're truncated from 14 to 10, even at 10 you have recorded/save more DR than you can in 8-bit (RAW or Compressed)

Although his focus is photography, Bill Claff has done amazing stuff: http://www.photonstophotos.net/Charts/PDR.htm

[Shirozina]

29 minutes ago, Andrew Reid said:

With Magic Lantern

They chop the bits off at certain ISOs, giving 8bit RAW

Go and test that and compare to the 14bit raw

Then you will see a direct example of the advantage of higher bit-depths like 8bit vs 10bit.

Yes in order to get the data rate down this is a necessary compromise. All other cameras inc the 5D3 when generating internal codecs will be using the full bit depth to start with and not truncating it before it goes through the internal image processing pipline where it is tonally compressed into an 8bit space and then data compressed to meet the recording medias write speed limits

 

27 minutes ago, Andrew Reid said:

Yes but that DR is at the expense of something else.... Which is what I am telling you guys all along!

With the limited bandwidth of the internal codec then obviously this is the case as evidenced by the 4.2.0 codec vs 4.2.2 - how Panasonic can get a 10bit 4.2.2 4k codec to work at  a pretty low 150mps is amazing.

10 minutes ago, maxotics said:

That's not how I understand it.  Again, I'm a hard-data guy.  8-bit is 8-bit is 8-bit   Larger sensor cameras can capture more theoretical DR, which is born out by DxO and Bill Claff tests.  There is more DR from a A7S, especially in low light, because its pixel to area ratio is much higher than MFT sensors with more pixels than are necessary for video (but important for photography).   

One of traps is assuming an 8-bit value from compressed video matches an 8-bit value from RAW.  There are NO real 8-bit color values in compressed videos.  They are pseudo-generated from a 24-bit color value which itself is saved in a 6 color matrix in some 4:X:X form.   In RAW, the 14bit value is a REAL value of a red, green or blue filtered pixel/sensel.  There are no real colors in RAW, or dynamic range.  They are just physical readings of light hitting silicon.   In sense, you HAVE to pick a black level in all 8-bit compressed video.  When shooting 10-bit RAW, you don't need to do that so you can be off a stop in your exposure and set your black level to make a nice image, whereas in 8-bit compressed video, you're more locked in.

It's physically impossible to capture more than 8 stops of DR if you only have an 8 bit signal from the A/D - if you are 'hard data guy' then this should be easy to understand

[maxotics]

9 minutes ago, Shirozina said:

It's physically impossible to capture more than 8 stops of DR if you only have an 8 bit signal from the A/D - if you are 'hard data guy' then this should be easy to understand

HA HA.  Couldn't agree more!  But I'd still rather have 8-bits RAW than 8-bits from a 4:2:2 data block   .... well, maybe, I've never tested that, not sure you could expose correctly, or don't need more than 8-bits of RAW data to calculate down to a decent image.  Another thing for Andrew to do

[Ilia3101]

24 minutes ago, Shirozina said:

It's physically impossible to capture more than 8 stops of DR if you only have an 8 bit signal from the A/D - if you are 'hard data guy' then this should be easy to understand

If it is linear data like RAW, then yes, you can only fit 8 stops mathematically, but in the case of gamma corrected data or LOG, the stops are spaced out more equally, in LOG they all get an equal amount of the range, it is LOGARITHMIC, so you could compress as many stops as you want in to the 8 bit 0-255 range, all the way up to 256 stops, if you want to give each stop a range of only 1.

 

EDIT: I was BSing, just reread your post and realised you mentioned 8 bit out of the A/D 

 

@maxotics, 8 bit RAW is likely to look much worse than 8 bit processed 4:2:2 images, as RAW will have the 8 stop dynamic limit due to being linear. Although 8 bit raw would be fine if there's lots of noise.

[Shirozina]

11 minutes ago, maxotics said:

HA HA.  Couldn't agree more!  But I'd still rather have 8-bits RAW than 8-bits from a 4:2:2 data block   .... well, maybe, I've never tested that, not sure you could expose correctly, or don't need more than 8-bits of RAW data to calculate down to a decent image.  Another thing for Andrew to do

 If you are happy with only 8 stops of DR then go with it on the 5d3 ML setting

I used to shoot slide film back in the day with less than 8 stops of DR - and get paid quite well for it!

[maxotics]

2 minutes ago, Shirozina said:

 If you are happy with only 8 stops of DR then go with it on the 5d3 ML setting

I shall resist a plunge into the abyss