Reconstructing blown highlights

darktable 4.0 user manual states that:

A better alternative [to using the highlight reconstruction module] is to disable highlight reconstruction and instead use the reconstruct tab on the filmic rgb module.

Watching Aurélien Pierre’s video https://youtu.be/F3CeJ1F4wV8, I had instead the impression that highlight reconstruction were more powerful.

Which one should I prefer as a beginner?

1 Like

The video is good but now dated… I would do the following but there will be other advice…

Use the new default inpaint/segmental highlights that will be applied. Use the white relative in filmic and then use the tone eq… to manage the highlights…if the issue is extreme then you can move to trying filmic reconstruction after that

V4.2 is going to behave differently in the default to your image wrt highlights. The previous default of clip highlights often led to cyan or magenta issues when used with v6 filmic and the default of rgbmax…

Hi, I don’t get it, sorry. With my darktable v. 4.0.1 I have highlight reconstruction on by default, set to clip highlights.

Yes well for some months there are now better options… you would need to build the code or use one of the many builds that are done nightly or weekly to access that…

I would say in 4.01 the best option if you are using v6 filmic and defaults and the modern wb vs legacy would likely be to disable HLR esp clip highlights and use filmic HLR with or without the guided LPHR as AP showed but you can mitigate it also by using some work arounds until the release of 4.2 in the next month or so…

Before the newer methods were in the code I would often stick to and still often do legacy wb and I would use v5 rather than v6 of filmic as I found it easier to manage with less potential artifact… if you recall in the video that it seems you watched… AP mentions how the existing methods are often not well suited to working with the modern wb using CC module…

All this is not really a problem in the current code…at least IMO

I think the manual is just wrong. I think AP wrote that, but it is not accurate in my opinion. Use GL or the new algos in 4.2 (in a month).

I think the premise for those comments is demonstrated in that video beginning around 24 min mark and the demonstration around 25:30 … Basically introducing the concept that the existing methods were not compatible with the modern wb using CC. At the time of the video all there was as an alternative would be filmic and GLP … so if you believe the case as laid out by AP in the video in that 5 or so min window starting at the 24 min mark then the comments would be appropriate likely but with the reworking of HLR and the new methods I think this statement and that text should be revisited…

Depending on the image, you may find that the highlight reconstruction module works fine by itself especially if set to something else than ‘clip highlights’.

Personally, I never had very much success with guided laplacians or the filmic reconstruction, but that’s mostly just because I didn’t fully understand it. AP in his videos certainly made it work… Guided laplacians can be very slow on my computer too - if the diameter is set to a high value, as it needs to be for large burnt out areas.

@solitone , I don’t know if you’re running windows or something else, but if you are on windows, the current weekly builds posted on this site seem very good, and they have all the features that will be present in dt 4.2.

The reason I mention this is because (IMO) highlight reconstruction could be difficult in dt 4.0, and there’s new options now in the highlight reconstruction module which are very good - Todd was talking about these above. Either way, it won’t be long till 4.2 is released now!

DT has multiple options for handling clipped highlights. The next stable release of DT will have some newer options as people have mentioned here. One way I have learned to handle highlights in DT is to use reconstruct in LCh and if there is a magenta problem with the highlights then reduce the clipping threshold until the magenta goes away. I personally rarely use the filmic reconstruction method. Just try the different methods and see which works best on the individual image. Good luck.

Ha Ha… so funny. I am now the opposite… I rarely ever used it…basically I didn’t really fully understand it but I will now if filmic itself , basic HLR ie inpaint or segmental, and most often I will get by with the tone eq…but if things are really blown I will go to it and you can get a nice result tweaking bloom and structure to blend things in… I must admit I have not really ever given the Laplacians more than a passing try…esp given how hard it hits performance… I have taken to using the tone eq a lot with single channel color blending to work and tweak highlights…

I agree for some images filmic can be very useful for highlight recovery. I just hate blown highlights as they give such grief to work with. Lch has worked well for me with images that only some channels are clipped.

That was me, to such an extent that I would always expose for the highlights to stop them from clipping. But in high dynamic range situations, that would mean the main subject is several stops underexposed, which means you get too much noise in post after boosting the exposure. And filmic doesn’t handle large gains in exposure well (especially v6 at max RGB).

So now, I’m trying to concentrate more on the subject and just let some highlights blow. It’s all about compromise of course. I really hope the new highlight tools in 4.2 can mean some graceful blown highlights, although last time I heard, the new tools were only for Bayer sensors. Is that still true?

I think both now…

I think 4.2 is pretty good, going by the dev versions I’ve been using. The new highlight recovery options from @hannoschwalm are just great. I rather think at least some of the options are available for X-trans sensors too now but not sure.
The new sigmoid module is pretty good at sort of smoothly fading to white in highlights if you let it too, a bit like filmic V6 no preserve but different.
Edit - just saw Todd beat me to it and with better info!

Sounds great. I usually install dev builds but been too busy with work since the last dt release, so this new release will be a nice surprise. I already have a few photos tagged to try with the new Sigmoid module because I’m not happy with the filmic versions.

A quick comparison… on a corner of a table in sunlight :slightly_smiling_face:
On the examples below I’ve dropped the exposure way down so there’s no interference from filmic or anything else.
This is using the very simple (to use) inpaint opposed, the new default. The other segmentation based has a lot more controls, which I haven’t quite fully got my head around yet but it works great too

No highlight reconstruction:

New default HLR inpaint opposed used with ‘modern’ wb (WB and Color calibration)

Same with legacy WB:

Old reconstruct in Lch with modern wb.


This one looks a bit better in legacy wb mode (i.e. no color calibration):

And the worst of all, the 4.0 default, clip highlights… in legacy WB:

This one is even worse in modern WB:


I haven’t included a guided laplacians example, because I find it hard to use - I think it can usually do a reasonable job but it slows my PC down to the point of being unusable - after each slider adjustment I might as well go and have a cup of coffee… OK not quite that bad but you get the idea :wink:

1 Like

Filmic plays better with the new HLR… it seems like the combination of the changes to v6 with the interplay of the older HLR methods and the modern WB methods created more often than one would like “scenario specific behaviours” … I like that description, I stole it from @MStraeten I believe he used it in a facebook post… :slight_smile: … Filmic defaults have been tweaked an the new HLR helps and sigmoid gives an option so I think overall people should be happy. I found before sigmoid came into the code that I reverted to v5 filmic with preservation set to no and used legacy wb and I had really few issues with the salmon colors or the cyan and magentas in highlights… So I think people will like the refinements added to 4.2… Snapshots can now be scrolled and panned in sync which can be nice for comparisons… lots of other new things but I think you will like the changes…

3 Likes

It is a beast but also I think AP did a better job in his later video of showing that it is most effective at least in the example he presented at reconstructing the underlying gradient in the blown areas and then you manage the reset with filmic… so I think people enabled it and maybe played with some slider and got very little out of it for the hit… but in its intended use and with specular highlights I think its quite good…having said that I rarely use it… and now with the new methods it might become even more niche overall…

1 Like

Yes, I was impressed with how he managed a Play Raw image using GLs in a more recent video… it’s just very hard for me to use! Maybe when I get my new GPU sorted it will be usable :smiley:

Ha I think the thing for me is that if my photo is that bad I am likely chucking it anyway…

1 Like

Lets not forget that Mr. AP’s first bit of advise is to use your camera’s exposure compensation button and prevent blown highllights in the first place.

2 Likes