I Have Published a HDR10 to HLG Converter

[kolak]

Originally Posted by wswartzendruber

Using Y for MaxCLL is outright irrational. Let's say the value is signaled at 593 nits. That can mean anything from 6% equal intensity across all three channels to 100% intensity on just blue.

No idea. I just read it in tool release notes.
6% on R,G,B gives bright white? I thought it's about black and Y will be low (at eg. 70 and 70 is not 593 nits for sure)?

[wswartzendruber]

Y(r = 0.06, g = 0.06, b = 0.06) = 593
y(r = 0.00, g = 0.00, b = 1.00) = 593

[kolak]

https://www.itu.int/dms_pubrec/itu-r...0-I!!PDF-E.pdf

table 4 has this:

Y'=0.2627R' + 0.6780G' + 0.0593B'

are they referring to it?

[wswartzendruber]

Originally Posted by kolak

https://www.itu.int/dms_pubrec/itu-r...0-I!!PDF-E.pdf

table 4 has this:

Y'=0.2627R' + 0.6780G' + 0.0593B'

are they referring to it?

Those are the BT.2020 coefficients, so yes. But you can apply them to display light as well.

[kolak]

So how do they count light then if they refer to this table?

[wswartzendruber]

Originally Posted by kolak

So how do they count light then if they refer to this table?

I have no idea. MaxCLL is specific to PQ, so how it's taken should reference BT.2100.

[kolak]

Have you implemented any filtering for compressed formats?
Like skip 1% of brightest pixels, low pass filter?
If you take properly 1K nit mastered master what MaxCLL is your tool reporting?

[wswartzendruber]

Originally Posted by kolak

Have you implemented any filtering for compressed formats?
Like skip 1% of brightest pixels, low pass filter?
If you take properly 1K nit mastered master what MaxCLL is your tool reporting?

I haven't done any filtering. hlg-tools includes a utility called pqstat, but it doesn't filter. You'd need to do that via FFmpeg or somesuch.

I have heard rumor that there are defined specifics on how this is to be done, but I don't know what they are.

[kolak]

There are none.
Each company does own based on tests.
Paper math is not designed for compressed sources.

Quick test v210 vs ProRes gives on some random short file max Y as 922 vs 926, so it would translate to few nits, but not that bad actually.

[wswartzendruber]

Originally Posted by kolak

There are none.
Each company does own based on tests.
Paper math is not designed for compressed sources.

Quick test v210 vs ProRes gives on some random short file max Y as 922 vs 926, so it would translate to few nits, but not that bad actually.

I guess I may as well start messing around.

[kolak]

Well this was ProResHQ. Now do the same with 30mbit h265 4:2:0 and things most likely will look way worse.
Maybe good solution is something which would average pixels so smallest one is made of 4 original. Who cares about place in the frame which is 1 pixel big? Maybe 4 is also too small?
Half size resize with area filter seems to give quite good (923) results on this sample
This is 30 seconds test, so don't rely on it

[wswartzendruber]

Originally Posted by kolak

Well this was ProResHQ. Now do the same with 30mbit h265 4:2:0 and things most likely will look way worse.
Maybe good solution is something which would average pixels so smallest one is made of 4 original. Who cares about place in the frame which is 1 pixel big? Maybe 4 is also too small?
Half size resize with area filter seems to give quite good (923) results on this sample
This is 30 seconds test, so don't rely on it

I'm going to start by having FFmpeg blur the living crap out of it.

Eh, not that much, though.

[kolak]

Half size seems to work
Not sure if blur is a good way.

Interesting fact. h265 performs better than ProRes. Maybe it's not a surprise actually as it's way more complex codec and handles contrast much better. 924 max value, so not that big deal. Looks too good actually
Maybe 20Mbit is a lot for HD.

[wswartzendruber]

Originally Posted by kolak

Half size seems to work
Not sure if blur is a good way.

Interesting fact. h265 performs better than ProRes. Maybe it's not a surprise actually as it's way more complex codec and handles contrast much better. 924 max value, so not that big deal. Looks too good actually
Maybe 20Mbit is a lot for HD.

Actually, yeah! Except I'm going to quarter-size 4K down to 540p.

EDIT: Quarter-size on each axis.

[kolak]

Yep, quarter size may be even better and faster to process (tried on HD as well and it was good). How much would it affect MaxFALL calculation?
Maybe I'm over-exaggerating it.
ProRes will do few levels overshoot, around 4-8 max (so far I've seen by simulating very high contrast edges), so it's not that much. How many nits is this in worse case?

5mbit h265 for HD is doing up to 20 Y levels, so this is bit high.
Wavelet based codecs (Cineform, Jpeg2000) behave bit better. DCT based always go up, where wavelet ones also bit down.

[wswartzendruber]

I just got from pqstat that Iron Man has a MaxCLL of 2,682 when metadata declares 826.

I'm going to try omitting the first and last several seconds in case there's some frame decoding error and artifacts are getting rendered.

[kolak]

Hmm...2,682 sounds bad. No one would master it to such a value. It doesn't sound 2K with overshoots, just random value, so something must be not right. Try some 20 min. portion- should be representative and give max as well (no guarantee of course).
Studios master mainly to 1K, 2K and 4K nits and typically grading software will clip max to this values (+- few nits). This is why value like 2682 is strange.
How old is this title? Try something from last 6 months.

Why can't we measure it this way.
All delivery codecs are YUV based, so just find max Y value and then count light by comparing it to PQ curve. Eg. I know that 769 (for 10bit) is 1000nits (this also perfectly agrees with Resolve when grading).
This sounds simple and should be correct (overshoots problems still apply though).
Why do we need some convoluted way to recreate RGB when original grading decision is bound to Y value inside your file?

[wswartzendruber]

Because MaxCLL is not properly calculated using Y values. It is properly calculated with max(R,G,B).

[FranceBB]

To be fair, I have a few ProRes HQ over here and I was checking the MaxCLL value by comparing it with the one sent via xml by the official provider.
For Harry Potter and the Chamber of Secrets, the official value provided via xml is 2856 Nits, however the one calculated by PQStats is 3204 which means that it's off by 348 nits due to compression overshooting (yes even if it's 900 GB worth of ProRes).

Originally Posted by kolak

Hmm...2,682 sounds bad. No one would master it to such a value.

For that particular title, sure, but I've seen movies mastered at really arbitrary nits depending on which studio did it.
For instance, this is a series of titles we received over the years along with an xml depicting the MaxCLL:

Goodfellas 247 Nits
The Dark Knight Returns 1399 Nits
The Good Liar 244 Nits
Richard Jewell 1213 Nits
Harry Potter and the Half-Blood Prince 1915 Nits
The Dark Knight 1002 Nits
Harry Potter and the Order of the Phoenix 1135 Nits
Harry Potter and the Deathly Hallows: Part I 2632 Nits
Harry Potter and the Deathly Hallows: Part II 2811 Nits
Zack Snyder's Justice League 835 Nits
Batman Begins 992 Nits
Harry Potter and the Philosopher's Stone 552 Nits
Harry Potter and the Goblet of Fire 2848 Nits
Harry Potter and the Prisoner of Azkaban 1838 Nits
The Lego Movie 1687 Nits
Godzilla 561 Nits
The Town 1339 Nits
Just Mercy 1017 Nits
Birds Of Prey 4616 Nits
Harry Potter and the Chamber of Secrets 2856 Nits
Dune 945 Nits

so as you can see, it looks like the MaxCLL can be as arbitrary as people like, so beware of what the distributor puts on the BD 'cause it might differ from the actual value of the master.

[kolak]

Sorry, but MaxFALL is something totally different than I thought.
Not sure why it’s defined this way, but it has not much to do with fact that you graded to eg. 1K nits. Grading can be 1K where MaxFALL 1.5K ( I assume). Is it true MaxFALL will be always bigger ( or at least equal ) than max brightness coming from Y?
What I see in Resolve and on any other scopes etc. is a separate thing to MaxFALL. Titles are graded to rather round values and it’s 1,2,4K etc. but MaxFALL is totally separate thing. Why, no idea.