[DT Master 4.1.0] color balance rgb perceptual briliance grading highlight slider creates artifacts

[s7habo]

Describe the bug/issue

If the highlights of the photo are close to the limit of dynamic range, the highlight slider of the perceptual briliance grading in the color balance rgb module produces artifacts when a certain threshold is exceeded.

This only happens in the default darktable USC (2022) color space. In JzAzBz (2021) color space the slider behaves as expected.

To Reproduce

1. increase the exposure of the photo with exposure module until the highlights are at the limit of the dynamic range.
2. in the color balance rgb module, increase the value of perceptual brilliance in highlights using the highlights slider.
3. After a certain value, the artifacts - wide white spots - appear (see screenshot/video below).

Screenshots

Screencast

Short demo:

CB_rgb_artefact.mp4

Platform
_Please fill as much information as possible in the list given below. Please state "unknown" where you do not know the answer and remove any sections that are not applicable _

• darktable version : 4.1.0~git272.5a1a1845-1
• OS : Linux
• Linux - Distro : Ubuntu Studio 22.04
• Memory : 31,3 GiB
• Graphics card : NVIDIA GeForce GTX 1070/PCIe/SSE2
• Graphics driver : 470.141.03
• OpenCL installed : yes
• OpenCL activated : yes
• Desktop : KDE-Plasma-Version: 5.24.6

Additional context

I can't say for sure, but I suspect that this problem may be related to the

#12233

since this behavior appeared after that commit.

[todd-prior]

I think AP generates this in the last 10 min of his most recent video and he comments that it is due to gamut mapping....

[Mark-64]

I have this too and I believe that it is generated in filmic, maybe by reconstruction kicking-in at a certain point when increasing exposure or brilliance.
Reconstruction is never really disabled, the threshold is set high instead, but it can alway be activated if exposure is set high enough

[pehar1]

@Mark-64 , this has already been discussed here : https://discuss.pixls.us/t/something-wrong-with-darktable-4-1-0-git272-5a1a1845-1-clipped-highlights/32540

[pass712]

Can reproduce under Windows.

dt: 4.1.0+283~g5cca9d64b
OS: Win10
Mem: 16 Gb
Graphics card: NVidia GEFORCE GTX 1060 6GB
OpenCL installed : yes
OpenCL activated : yes

[jenshannoschwalm]

As @s7habo hinted the problem is somewhat known and discussed in the mentioned pr.
Pinging @flannelhead here...

[flannelhead]

The fix that was done in #12233 was about avoiding divisions by zero.

Here's the result with current master:

With that PR reverted:

Notice that in the latter image there are specular highlights with black pixels. That's what the fix meant to address. However, it seems here that neither result is really acceptable.

As others have concluded here, the large blotches are caused by filmic's reconstruction which seems to spread those specular highlights. color balance rgb's brilliance adjustment in dt UCS mode seems to push specular highlights quite far, so filmic's reconstruction picks up these very bright highlights and spreads them all around.

Contrary to what some stated about filmic's reconstruction, at least in the case of this image it can be practically disabled. Turn the threshold all the way up to 6 EV and transition slider all the way down to 0.25. At least with those settings I could no longer reproduce this sort of artefacts.

It might be worth investigating why color balance rgb brilliance adjustment causes those insanely high values for specular highlights. However, the fix in #12233 doesn't cause them in itself, I believe. It just makes the issue more visible due to triggering filmic's reconstruction into generating those blotches.

[jenshannoschwalm]

Would there be a way to check in either module for such an issue and give a "module warning"? (BTW I find the black pixels far worse as "not so easy to catch" if not pixelpeeping.)

[flannelhead]

BTW I find the black pixels far worse as "not so easy to catch" if not pixelpeeping.

In the case of this image, yes. But it can happen in vastly larger areas (e.g. sun disc) so that's a no-no anyway (in my opinion).

Would there be a way to check in either module for such an issue and give a "module warning"?

I'm not sure. The best thing to do would be probably to fix the brilliance formula (or clamping) in such a way that avoids those vast changes of bright highlights and clamps them to some sensible range. That was already briefly discussed in #12163 but would need more investigation because there didn't seem to be a viable fix yet.

[todd-prior]

I dont know if it offers any clues but the artifacts do not appear the same at different zooms in the central preview and dont match the preview used for panning and zoom navigation which seems to briefly show during panning on the center display and then disappears and this also is nothing like what gets exported... I checked and got the same with or without Opencl. I posted on the topic with a video and still example..... For sure the strongest impact seems to be the filmic transition slider but the data used being different might indicate more???

[herrkami]

I just edited around 200 images and encountered the problem in around 20 of those. In five cases the blobs revealed themself only after export, because of the zoom level dependency.

[da-phil]

I can also reproduce this issue on a couple of my images too. The "workaround" described by @flannelhead works though.

[TurboGit]

@flannelhead : Do you think you'll have time to handle this for 4.2?