I was just wondering if anyone is having an improved experienced using the guided laplacians method for highlight reconstruction? In my hands it is very resource intensive and gives disappointing results. Maybe other users are having more success and can share their experiences with this option.
2 Likes
I have tried it a couple of times but I donāt really get anything predictable. I always blame me and not understanding somethingā¦ It would be nice to see a few success stories and results to confirm expectations. I think AP did mention it in one of his videos but I think I found it hard to gauge the results on screen from thatā¦ again I totally blame me at this pointā¦
I am the same. I blame myself and hence why I worded my question the way I did. I donāt expect AP to create something that does not work.
Well I have tried it a few times on skiesā¦but reading the limited documentationā¦it seems like it may really be for specular highlight and candles and streetlights etcā¦again just guessingā¦ maybe this is why I really didnāt notice any thingā¦
" Use an algorithm (derived from the diffuse or sharpen module) to replicate details from valid channels into clipped channels and to propagate color gradients from valid surrounding regions into clipped regions. This is a slow and computationally-intensive method designed for maximum smoothness and seamless blending of the reconstructed regions into their neighborhood, and is designed primarily to reconstruct spotlights and specular reflections. This mode is available for Bayer sensors only."
I find one of the better ways to work clipped highlights is to first explore the available methods in the highlight reconstruction module. I am working an image at this moment where I have chosen reconstruct in LCh to recover the details. I also experiment with lowering the clipping threshold despite the warning that we should never have to use this. I find it can be very helpful to move it to obtain a desirable result.
I try to get what I need from filmic HR but I have done what you mention in your post as well and even use the color reconstruction module here and there. I have recently seen somewhere that Aurelien was suggesting new defaults for filmic, contrast 1 setting the curve toe and highlight sections to hard , latitude at zero and starting with exposure at 0.7EV. This got me just tonight experimenting with this and also the hardness setting which is on auto but you can get quite a different curve if you play with it and the curved section settings and it impacts highlights a fair bitā¦ so after all this time still playing around with filmic to see what I can get with it in certain situationsā¦
What I love about Darktable is the multiple approaches that can be used. That is probably what some people hate about Darktable. I describe Adobeās Lightroom to my students as the automatic Hyundai that is great to drive around town or to the shops for groceries, but Darktable is the Lamborghini and it even lets you reconfigure the engine and the gearbox if you want.
2 Likes
I had the same problems using this method. After trying different situations I could get acceptable results under the following conditions:
- damaged area is relatively small compared to the whole image.
- damaged area is part of a more or less evenly coloured surface so that the colour propagation of this reconstruction method leads to reasonable results
Even in this cases it is extremely resource intensive (very slow) and I guess it will be a rarely used option in my workflow. Good to have it anyway.
It is difficult if not impossible to reconstruct large portions of the image (sky, walls, etc.) especially if there is no suitable colour next to or surrounding the damaged area.
1 Like
@hannoschwalm is also developing another highlight reconstruction method based on @Iainās work in gmic. Not sure when thatās supposed to be done, but it looks like itāll be better at reconstructing large chunks of blowout.
1 Like
Tried it and I honestly donāt understand it. Iām getting similar or worse-than-before results on every image Iāve tried it on. Perhaps itās for specular highlights only. Iām hoping a video by AureliĆ©n could shed some light on how itās supposed to be used.
What Iām really hoping for is @hannoschwalmās and @Iainās work. In the thread where guided laplacians and their method were compared I think it looked considerably better.
This is what the documentation indicates, specular and smaller parts of the photo.
1 Like
I had the same experience: small blown-out regions are reconstructed quite well. Sometimes I had to use diameter of reconstruction = 2048 px (insanely slow); but very often, the default of 128 px or perhaps 256 px works better (AurĆ©lien did warn about propagating false colours). Same for iterations ā anything between 5 and the max, 64.
1 Like
Well, āhighlight reconstructionā is and will always be ābest guess onlyā : you canāt recreate the information thatās lost, only try to fill in the holes in a reasonable way. The larger the hole, the harder that getsā¦
Thereās probably also a difference between āall channels clippedā and "one still has signalā.
So I consider any highlight reconstruction as an attempt to salvage whatās basically a lost cause (yes, I do use it on my images when needed :P). That means Iām not disappointed when I donāt get a perfect result, and quite happy when it does work 
.
2 Likes
Thereās a plenty of such insight in this video: https://youtu.be/34iZotlYBBs
2 Likes
Using the highlight reconstruction module I have been really impressed with the level of detail I was able to recover in some images when not all channels are clipped. I photographed a breaching humpback whale and its white underbelly was clipped. Highlight recovery module was able to recover the striations along the belly. If all channels are clipped I suspect we are in real trouble. It is obviously best to avoid clipping highlights in the first place, but that is not always possible. Highlight recovery seems one of the biggest challenges for all software.
Thanks @flannelhead!