Have an issue with a transition in a highlight area of an image. I am newish to Darktable, so I would like to ask for some guidance.
The image is a holiday snapshot with bright, but easily recoverable sky. In Capture One the transition is perfectly smooth, same in Raw Therapee. Darktable produces a clear division line going across the sky portion. It is there even without any highlight treatment/recovery. Easily demonstrable by simply pulling down exposure slider.
_DSC0060.NEF (55.3 MB)
I’m on darktable 4.6.1 and I can’t reproduce what you are referring to. Can you please show where in the photo there is a division line?
Have a look at the screenshot above. I pulled down exposure slider to bring it up. It goes from the corner of the roof diagonally down to the right
Are you on 4.6.1?
Yes I am, 4.6.1
Your image is overexposed in that area. The raw processor is trying to recover the area.
1 Like
In filmic set your highlights reconstruction threshold to about -1. Seems to handle it nicely.
@stuntflyer I did, big improvement, still there though, just…
@Dontask Yes, that gave me the best result, but any further changes re-introduce the transition line.
I am trying to end up with an image looking more or less like this one from Capture One. While following your suggestions I was able to smooth out the division line, any further changes bring it back. Exposing for highlights is probably the best fix here 
are you using masking or tone equalizer?
That’s exactly one of the cases where the simple opposed algorithm fails at exactly the transition from nonclipped-to-clipped data. You should use “segmentation based” here.
5 Likes
@paperdigits Mask over the sky area. Within the mask I tried filmic RGB, tone equaliser and both in combination. Couldn’t completely remove the transition.
@hannoschwalm perfect, this did the trick. Can manipulate the area further now, without transition line coming back.
Thanks everyone for your help, much appreciated
3 Likes
@hannoschwalm I see for my self that segmentation based works wonders on this image. I notice the default is inpaint opposed and that usually does a great job, but would segmentation based be a better default or is it just image dependent?
1 Like
There was another image that really showed this as well back when you were first working on it. I think maybe it was an image provided or introduced by @kofa . It was a statue in a garden against a bright background if I recall but it was the same the segmentation approach was perfect whereas the other methods failed…
In paint is simpler to use and works for most images. If you make the segmentation one default, then we will have complaints about too many slides to adjust.
Inpaint opposed is much (4 times here) faster than segmentation and in most cases (like speculars or small parts) “good enough”. So a good choice as default.
The tricky part for HLR algos is “how to guess a missing signal”. In inpaint we calculate some “correction data” from the whole image at the borders from unclipped to clipped per RGB color. This is mislead for images with very smooth transitions like here (in the sky).
The segmentation part of the algo does not do this via some global statistics but a) does a segmentation of clipped areas (per color channel) and for every segment chooses a “best candidate” to define that “correction data” per segment! That’s why the transitions are so smooth.
5 Likes
Hanno has explained the masks of the segmentation-based method here (you may want to also read my questions in the comment above the explanation):
Good info there, thanks for pointing it out