Originally Posted by fhoech
Sure, I'm just saying that this is not a conclusive test as to where in the chain the clipping occurs (i.e. if scientific looks like it clips it might just be because it sends more saturated values which would still show detail if the display had an actual 2020 gamut). For a real conclusive test the display shouldn't do any extraneous processing on the signal (not even P3-within-2020 to actual display mapping, which it has to do if it receives 2020 data).
I just found it odd that there is any visible clipping at all in your case (if I set a 2020 target in madVR with the pixel shader parameters you're using, I'm not really seeing that, but then again, I can only look at the clips that were shared, not at any scenes outside of that). Maybe 'clipping' isn't the right word and you're referring to a certain 'flatness' that may be in some of the rolled-off areas, I'm not sure.
I've run a few quick tests to rule out a few options.
First I have exactly the same behaviour when I disable the 3D LUT, so that's ruled out.
Second I tried a quick DCI-P3 calibration, using my DCI-P3 profile with a 2.4 gamma and telling MadVR the display was already calibrated to P3 2.4, I didn't have the time to run a 3D LUT). I had exactly the same results.
Looking closely at the picture, I think what happens at 500nits is that dumb has red but also yeallow/orange in the shot, while scientific replaces all the orange/yellow with red, at very similar saturation/luminance as the red already in the shot. This looks like clipping, but I think it's more a case of "same red" than clipping per se at the edge of the gamut. As I said earlier, it doesn't necessarily happen on the most saturated/brightest reds in the shot, it simply happens to be more red replacing the orange/yellow showing on dumb. So red on red = red everywhere, hence less detail than in dumb.
If I use 600nits target, there is more headroom for the highlights and scientific shows more levels of red, which means there is less clipping (if any).
Does it make sense?
Again, that shot *should* probably be shown with a 600nits target, as it's a 4,000nits content going up to 10,000nits. A brightness factor of 5 (diffuse white of 20) is what I would normally use for that type of content with my 120nits display.
But for some reasons, as has always been the case, dumb/naive is less prone to clipping, which allows to lower the target to 500nits while keeping enough details in the highlights, resulting if far better results in darker titles, such as the infamous 6nits shot in The Revenant.
I'll need to do more tests with other 4,000nits titles, to see if this "red on red clipping" also happens with other 4,000nits+ titles. The Shallows and Secret Life of Pets are not good examples because there are very few explosions/fires at 4,000nits+, given that they are 2,500/1,000nits titles.
The Dark Knight would have been great but MaxCll is barely above 1000nits, so none of these issues would show.
I tried with BvsS quickly and I have exactly the same issue. For example, look at 02:28:16, around the eyes, the red clips (detail is lost) because the orange is replaced by red and red on red = less detail. On that shot in BvsS, there is clipping (loss of detail) even at 600nits in scientific compared to dumb. At 500nits it gets worse and it adds ugly contouring. Dumb keeps a lot more detail (and no contouring) even at 500nits.
As soon as we get profiles allowing us to test for content max brightness, this won't be a problem anymore as we'll be able to swap profiles according to content.
In the meantime, as I don't fancy changing things manually for each title, I prefer to use a compromise that works well for all titles, and that's dumb convert gamut late with a target of 500nits. I actually enjoy the overall boost to the picture brightness with a brightness factor of 4 (diffuse white of 25) vs my usual 5 (diffuse white at 20).