#apertus IRC Channel Logs

that some of the luminance values coming of the sensor will not be linear?

intracube: Especially those near the saturated end (the full well region) as well as those near the bottom (near the noise floor)

(In a majority of cameras. For example, if you have a DSLR kicking around, pop a photo of skin near a hot lamp. The image will often look almost posterized because the internal camera matrix is not holding up near the nonlinear roof. Worse, the moment that a colour axis deviates from the learned red / blue, our perception will flag it immediately.)

intracube: Senese?

troy_s: sorry, I wasn't following channel

troy_s: yep, got it

hmm, I was under the impression you'd get that effect even with a theoretically perfect sensor (on an object that isn't white)

effect is often visible on sunrises, or as you say peoples faces

all three channels would clip to white at the same point (in a theoretically perfect sensor)

intracube: You've seen some handling for tonemapping and some of the implementations use a 'preserve saturation' which is actually somewhat counter to the aesthetic learned through film.

intracube: If you want to see whack, simply try blender.

intracube: You actually sort of want a 3D LUT that converges based upon some threshold.

intracube: So say red gets above 95% or whatever you want, you actually want the other channels to come up to desaturate toward the neutral axis.

intracube: Because the magic of film was that light nuked the three layers and you actually desaturated on your way to white.

intracube: If you think of what happens in Blender (or any default agnostic hard clip LUT)

(and avoid overexposing when shooting)

troy_s: so you're suggesting the Beta camera should be set up to roll highlights up to converge to 'white'

intracube: No.

so a post production procecss?

*process

exactly what I thought :)

intracube: You want accuracy. Because in post, say you have three cameras, you want to match footage etc. for CGI or even simple grading.

absolutely

You can ask anyone that worked on the Hobbit.

So they had to accurately arrive at it (via math)

That means these sorts of issues are paramount.

Same goes for some default massaged imaging LUT.

intracube: The problem with graphs and charts is that they don't really give you the idea of response in a random context with regard to the standard observer.

intracube: It's crazy. Don't forget, it is _not a spectral sensor_. That is, light sources, refractions, reflections will change that response.

(Mostly Rat Piss Yellow off of foreheads.)

so many graders seem to love the look tho ;P

After a transform, you'll end up with negatives in that noise floor, and clamping actually increases the stop of red.

hmm, yep

intracube: https://www.reddit.com/r/photography/comments/3q8q7o/purple_noise_in_shadows/cwds00b

see the linked thread, at https://www.reddit.com/r/photography/comments/3q4tnz/how_to_correctly_push_5_stops_with_canon/cwcbyb5

I am also reasonably sure that both explanations aren't quite as accurate.

The bottom line is that even if the calc were in float, the results would still derive negatives due to the transform.

Offset isn't an issue.

yes, offset actually helps (if you do noise reduction or frame stacking, you really need that negative data)

Matrices suck

They work wonders on idealized colour spaces and transforms, but for real world nonlinear things, they simply aren't effective.

agree, and I've noticed they fail miserably on highly saturated colors

and that happens with perfectly linear sensor data

Sensor data isn't perfectly linear.

None of them are.

Not really

That is a hardware thing. I think Bert could come up with a reasonable test to try and linearize the nonlinear regions, but I significantly doubt the accuracy of it.

(As in a merged 3D LUT that properly models the nonlinear regions.)

I am just skeptical of a technique that would attempt to attain a purely radiometric linearization of sensor data.

Colorimetrically.

(Or better, a light box and a spectro)

irieger: I think a Jeti with a good light box would go a long, long way.

https://github.com/ampas/aces-dev/blob/master/transforms/ctl/idt/vendorSupplied/arri/alexa/AlexaParameters-2013-Nov-13/Alexa_aces_matrix.txt

https://github.com/ampas/aces-dev/tree/master/transforms/ctl/idt/vendorSupplied/arri/alexa/v3/EI800

troy_s: in Alexa_aces_matrix.txt, CCT is color temperature?

DNG is similar, but with only two matrices

alexML: DNG is a joke.

alexML: But much like CMYK, the variables all sum up to unique transforms per set of contexts. So for the Alexa it is EI + [ND | no ND] + CA

irieger: Of course it does.

The gain ultimately shifts the matrix so it "does" I suppose is the best phrase.

So it might be deeper.

Bonkers.

alexML: Ask yourself how many vendors actually use DNG.

For good reason.

Adding overhead that is really not all that useful.

Which will let you commence work.

DPX already has enough metadata for colorimetry if you desperately want it, or even EXR.

troy_s: Quiet a few people seem to use it with some Sony Cameras and the Convergent Design Odysee 7Q rekorder. And Blackmagic

irieger: The essence of my point still stands however; it is an intermediate and as intermediates stand, there exists another with longer tails.

It is just a silly discussion in my eyes.

Sure. I know, the big three have their own formats. But for what I have experience with DNG works as good or bad as the other vendor formats in Resolve for example. Only thing is that only a small number of color spaces is supported by Blackmagic as the target after debayering

irieger: Bear in mind.

troy_s: And if it works good, why should I go to exr first? Compressed DNG is much smaller than EXR (with PIZ or so) due to having only a bayer image.

So two facets: A company with a vested interest DNG writes software _and_ the fact that Resolve is for grading, which is the absolute last possible thing you do in a post production pipe.

It is scene linear and your data would go there once, stay there forever until the very tail

(Not that storage is ever much of an issue)

troy_s: Sure. But in most workflows they may go through a grading system before to get everything through one pipeline and one standardized process to exr before later processing ;-)

irieger: See that ACES repo? Tell me where it goes through a grading system?

Most folks don't get that grading is the last possible concern. It really is. It is a tremendous one, but in terms of pipelines, it is well after editing and the rest of post.

and the ACES repo is just the toolset to use in the system you like ;-)

Baselight and Resolve stations are incredible. But again, grading.

Typical pipe is something along the lines of:

The ACES pipe would help you get the raw archival frames into a uniform colour space, with ACEScg and ACEScc (linear and nonlinear compatible) references for CGI and Color Correcting respectively.

I wouldn't save anything in ACEScc ... Just stay linear for the files. But Grading often starts before or while shooting.

One lights for sure

A one light / quickie grade can even be used as an entry point for a final grade, but that isn't touching the frame data.

It's a look.

Did I say anything like toughing the data?

I was trying to say that the idea of a one light has nothing to do with post work

It is just temporary work.

Anyways...

ACEScc is a damn smart grading space. It is basically a well researched set of primaries.

I love grading.

troy_s: I love ACEScc, did some stuff in it

irieger: Many folks don't get why manipulations done against primaries are important to consider the base primaries.

(Which offers solid results. DCI-P3 is another space that apparently offers decent results.)

Resolve 12 Colorspaces are pretty fucked up and pissed me when I first saw what they did

Resolve was / is designed for professionals at one point.

It is for the masses now I'd say. Just keep it simple at too much points

(Prosumer is really just consumer.)

They have some hidden stuff for the big studios, I'm sure they get settings help for the "Advanced" tab in the config which is nowhere documentet but is there internal setting testbet also

Yep.

Hints at a larger process. Larger pipeline. Bigger things.

You can't.

God knows I have watched this unfold over the (sad now decade+) time I have learned about them.

Just trying to figure out why a wide gamut display is different from a small gamut was a huge leap of cognition.

(I still suggest that understanding RGB as a _relative_ system is the entry point to understanding all of colour.)

Well "linear" is a nasty piece of work.

Language is huge here.

Scene linear and display linear, another bit of confusion.

Where scene linear will hold radiometrically linear values, and display linear is "only radiometrically linear up to a certain point then it doesn't really mean much"

I would love to hear your idea.

I think baselight is pretty straight forward. You have to learn about color management anyways and there I'd say they got a pretty usable concept

Even a basic scene linear to whatever transform is a nightmare.

You heard Huxley's quote on that right?

Your quote exactly

“The Will to Order can make tyrants out of those who merely aspire to clear up a mess.”

As we do research about HDR at our school I asked Peter at IBC/Blackmagic Booth: What color space is used when setting HDR PQ 1000nits as the output Color Space? P3 or Rec 2020?

I bet he stared.

Guess it?

All of the Academy folks have massive experience and deep math.

With real-world experience.

You are surprised?

nope. But hey, why don't they even document it?

Did you read Sayre's post on it regarding non-orthogonal basis vectors (discussion of why primaries matter during manipulation)

(Rick Sayre is Pixar's CTO for those not in the know. Ridiculous knowledge and comprehension.)

Very cool.

irieger: I love the whole motion picture pipe because it forces various disciplines to mix, and the result is a much better understanding of models and why.

We sat there as 4 people who worked on this HDR stuff in a presentation at IBC and when they showed a photo from their first tests we recognised the footage directly ;-)

irieger: You will be able to understand the whole thread https://groups.google.com/forum/m/#!msg/academyaces/9b4VuqPcOHQ/Do6yMSE3uZsJ

Thomas is _deadly_ clever too.

irieger: As a baseline, whorfin is Sayre.

Ok, good to know

Or some more sensortuning at first maybe

What are you on tomorrow?

Night irieger. Looking forward to your return.

irieger: Keep in mind...

(I believe they are axis aligned.)

I know there are differences

irieger: Do you use the CTL at your school?

(If not, perhaps you can begin integrating it.)

(shot on 6 or 7 cameras, VFX project)

i know

(The progress)

https://github.com/irieger/aces-dev / https://github.com/irieger/CTL

did some small hacks already. but as I said, hacks

postpro

Starting or knee deep?

just helped a bit with some small stuff. (Said they should go full ACES months ago btw.)

(It generates RGB all colour patterns etc.)

my ctlrender can do it already so why bother? ;-)

off to bed