How much bit depth? (10-bit / 12-bit / 14-bit)

[Mark Romero 2]

@kidzrevil Ah.. ok, thanks.

[kye]

7 hours ago, kidzrevil said:

I only shoot 14 bit. The 2-4 bits you chop off has a massive amount of data in them. The most I’ll do is convert the 14 bit file to a 12 bit cdng

Where/how do you see the data in those 2-4bits?  I mean, is it highlights, shadows, saturation, 'thickness' of image.. ??  I compared bit depths and couldn't see any difference, but I also don't know where to look

[tomekk]

6 hours ago, kye said:

Where/how do you see the data in those 2-4bits?  I mean, is it highlights, shadows, saturation, 'thickness' of image.. ??  I compared bit depths and couldn't see any difference, but I also don't know where to look

Current consensus is that there is a difference between 10 bit and 12/14, but not so much between 12bits and 14 bits. The last 2 bits are mainly noise, but I cannot find the exact quote from a1ex right now (main ML developer).

http://www.eoshd.com/comments/topic/21004-12-or-10-bit-raw-magic-lantern/?page=3#comment-168058 - 10 bit vs 14bit after lifting shadows

@kidzrevil - I would love to see a difference between 12 and 14 bits. Could you post your findings in this thread by sharing a short clip that shows the difference and, if possible, a few DNGs?

 

[kidzrevil]

12 vs 14 bit is visually identical but technically there is a difference between the amount of shades in a 12 vs 14bit file. You don’t really need a DNG for proof of that. The point of shooting in 14 bit is to avoid trunctuation or altering of the data present in the raw data when converting from mlv to cdng. @tomekk

[tomekk]

@kidzrevil You are right, but they cut out bits that contain mainly noise, where there are no visible shades anyway.

[kidzrevil]

18 minutes ago, tomekk said:

@kidzrevil You are right, but they cut out bits that contain mainly noise, where there are no visible shades anyway.

This is from the developer of raw2cdng. Its a visual of the difference between bit depths etc. the 2 bits of noise are very useful for generating noise profiles and still contain a significant amount if data over 12 bit. They are just visually imperceptible from eachother so you can get away with it

[tomekk]

58 minutes ago, kidzrevil said:

This is from the developer of raw2cdng. Its a visual of the difference between bit depths etc. the 2 bits of noise are very useful for generating noise profiles

It is a very good visualisation of how much data difference in total there is, but if I understand it correctly, ML is just cutting out the darkest blacks which contain mostly noise and does not re-distribute shades in higher stops of light. 

If you can use noise bits to your advantage, of course, use 14 bits. 

[cpc]

FYI, there is no reason to ever use any of the "maximized" variants in raw2cdng. 

10-bit raw can be all you need, if it isn't linear. ML raw is linear though, and 10-bit is noticeably worse than 12- and 14-bit  even though it is better than, for example, 10-bit raw from DJI cameras. 12-bit is actually quite good.

[kidzrevil]

@tomekk 12888 more shades of tonality over 12 bit is a hell of a lot of data. Whether most of that data is noise or not is irrelevant. The more tonality you have in post the better. Its the equivalent of 12 bit vs 14 bit raw photos

6 minutes ago, cpc said:

FYI, there is no reason to ever use any of the "maximized" variants in raw2cdng. 

10-bit raw can be all you need, if it isn't linear. ML raw is linear though, and 10-bit is noticeably worse than 12- and 14-bit  even though it is better than, for example, 10-bit raw from DJI cameras. 12-bit is actually quite good.

12 bit is very good. I’d still rather work with the true 14 bit image in post though. 12 bit is perfect for longer takes and I’d just avoid 10bit all together. Another advantage of 14 bit is it syncs the sound perfectly with raw2cdng

[tupp]

Not sure about that raw2cdng chart.  What is going on with the "un-maximized" 16bit cdng?

 

Almost every digital system (video, audio, whatever) "maximizes" the bit increments within the given amplitude range.  In other words, the maximum bit increment occurs at (or near) the maximum amplitude and the minimum bit increment occurs at (or near) zero amplitude.   So, the proportional range of bit increments in noise should be the same across all bit depths mapped to the same dynamic range (and to the same mapping curve).

 

Sure, mapping 16bit to the same dynamic range will put a greater number of shades into the noise, but that number of noise shades will be proportionally the same as the number of 8bit shades in noise with the same given dynamic range.

[kidzrevil]

@tupp the programmer labelled the 16bit file maximized because he stretched the values to fit the whole 16bit container. This prevented the “pink highlight” issue in adobe premiere. The unmaximized version are just the regular Canon values in the 16bit cdng which is interpreted correctly in davinci but incorrectly in premiere.

[tomekk]

30 minutes ago, kidzrevil said:

@tomekk 12888 more shades of tonality over 12 bit is a hell of a lot of data. Whether most of that data is noise or not is irrelevant. The more tonality you have in post the better. Its the equivalent of 12 bit vs 14 bit raw photos

 

It is a little bit confusing, but what ML does is not equivalent of 12 bit vs 14 bit raw photos. Shades are mapped out differently in "true" 12 bit vs 14 bit raw photos. ML just truncates bottom bits and does not change values in higher stops, AFAIK. 

Anyway, a theory is just a theory. Check in real life. If you see the difference in grading, that is great, but it would be nice to see those clips/dngs too. 

[tupp]

2 minutes ago, kidzrevil said:

@tupp the programmer labelled the 16bit file maximized because he stretched the values to fit the whole 16bit container. This prevented the “pink highlight” issue in adobe premiere. The unmaximized version are just the regular Canon values in the 16bit cdng 

Thanks for the explanation.

 

However, I am not sure how the "un-stretched" situation can occur.  Is the programmer taking a 14bit file and converting it to 16bit?

[kidzrevil]

26 minutes ago, tomekk said:

It is a little bit confusing, but what ML does is not equivalent of 12 bit vs 14 bit raw photos. Shades are mapped out differently in "true" 12 bit vs 14 bit raw photos. ML just truncates bottom bits and does not change values in higher stops, AFAIK. 

Anyway, a theory is just a theory. Check in real life. If you see the difference in grading, that is great, but it would be nice to see those clips/dngs too. 

yeah to be honest with you  I wouldn't even stress over it. The difference is visually imperceptible. One of the key reasons I shoot with it is the fact that my mlv converter seems to only sync the audio right with the regular old 14 bit hack. Everything else my audio is out of sync. If audio is not a thing for you then 12 bit is more than sufficient. I would be nitpicking if I could say I could spot the difference even next to each other. I just don't trust my mlv converter or the ml hack with these experimental builds.I can't afford to get a corrupted shot on a video for a client so my paranoia just wouldn't let me shoot in anything but the vanilla 14 bit lol 

26 minutes ago, tupp said:

Thanks for the explanation.

 

However, I am not sure how the "un-stretched" situation can occur.  Is the programmer taking a 14bit file and converting it to 16bit?

no he is putting the original canon value range inside of a 16 bit container since there is no such thing as a 14 bit dng. There are onlt 12 bit and 16 bit dng's and for the sake of preserving the original 14 bit values they are placed inside of a 16 bit cdng

[cpc]

1 hour ago, kidzrevil said:

no he is putting the original canon value range inside of a 16 bit container since there is no such thing as a 14 bit dng. There are onlt 12 bit and 16 bit dng's and for the sake of preserving the original 14 bit values they are placed inside of a 16 bit cdng

Premiere has issues with missing metadata. It doesn't care if the values are stretched. What happens is that it can infer missing metadata correctly when the values are shifted and zero padded to 16 bits. Also, there is 14-bit CinemaDNG, but Premiere has trouble with 14-bit uncompressed (not with compressed though!). Now that all ML dng is compressed, it should work fine with Premiere at 14 bits.

[kidzrevil]

@cpc thanks for correcting me ! Didn't know that was the cause of all the weird behavior of the files in Adobe. I wonder if someone could make a MLV converter that will convert to 14 bit dng's. Something to preserve the entire bit depth of canon raw without having to use 16 bit cinema dng

[webrunner5]

On 4/6/2018 at 12:29 AM, kye said:

Not sure if you're on the CML?  https://cinematography.net  I've only dipped my toes in the water a little there, but it seems to have lots of people who casually talk about RAW workflows for Red and Arri cameras.  A very different world than the one I live in!

Here is a test on there between an Arri Alexa and Film. You can see why so many people shoot on a Arri. Pretty hard to tell apart.

https://cinematography.net/Skin/skin.html

[Deadcode]

12 hours ago, kidzrevil said:

 The difference is visually imperceptible. One of the key reasons I shoot with it is the fact that my mlv converter seems to only sync the audio right with the regular old 14 bit hack. Everything else my audio is out of sync. If audio is not a thing for you then 12 bit is more than sufficient. I would be nitpicking if I could say I could spot the difference even next to each other. I just don't trust my mlv converter or the ml hack with these experimental builds.I can't afford to get a corrupted shot on a video for a client so my paranoia just wouldn't let me shoot in anything but the vanilla 14 bit lol 

no he is putting the original canon value range inside of a 16 bit container since there is no such thing as a 14 bit dng. There are onlt 12 bit and 16 bit dng's and for the sake of preserving the original 14 bit values they are placed inside of a 16 bit cdng

Here is an example for ML 14 bit:

1111 1111 1111 11

This is ML 12 bit:

1111 1111 1111 

This is ML 12 bit after you convert it to cdng: 

1111 1111 1111 00

The last two digit can represent 2^2 = 4 shades. This is the difference between 14 bit vs 12 bit recording with ML RAW.

With 10 bit recording the difference is 2^4 = 16 shades vs 14 bit recording with ML RAW.

 

[KnightsFan]

@Deadcode each bit you add doubles the number of possible values. Adding two bits means there are 4 TIMES as many shades. 10 bit has 1/16 the shades vs 14 bit.

[kidzrevil]

Yeah those extra 2 bits are massive because they multiply exponentially @KnightsFan