[Feature]: Highlight reconstruction for RAW files

[nicokaiser]

NegPy version / build

0.58

OS

macOS 26

Scanner

Digital camera (macro/copy setup)

Scanner model / details

Sony A7 + Nikkor 55mm f/3.5

Light source

CIneStill CS Lite

Correction mode

Off

Input ICC profile (if any)

No response

Relevant process-tab settings

No response

What happened

When clipping occurs in RAW files, clipped highlights may appear as pink, because of the Bayer sensor and the way white balance is handled.
Although this is an edge case and of course it is preferable to avoid clipping, it is sometimes not really avoidable, like when scanning slides with very thin parts (like the sun).

It would be great if clipped highlights would be reconstructed in this case. Darktable uses the "inpaint opposed" method, which makes highlights clipt to true white instead of pink:
https://docs.darktable.org/usermanual/5.6/en/module-reference/processing-modules/highlight-reconstruction/

I'll attach an example scan (which is not a masterpiece of photography, but shows the case) and its NegPy interpretation.

Source file

2017-04-09_Dia_Film2017015_012.ARW.zip

Screenshots

No response

[marcinz606]

try the demosaic options that were recently added, it was reported some time ago (.arw highlithgs turning puprple) and it turns out just using different demosaic algo gets rid of that

[nicokaiser]

I just tried this, but none of the demosaic algorithms did anything to the pink highlights for the file attached.

[thetalkingdrum]

Yeah, this is a know problem for camera raws and related to WB and not demosaic I believe . Should be possible to solve for positives. When it comes to negatives I think it is probably not a good idea. Here the correction would land in the shadow area = filmbase. So you would change the color of the orange mask, and shift the black point in some semi arbitrary way. I think for negatives there is anyway no valid reason for clipping like there is for slides.

I will take a look at how difficult it would be to implement.

[thetalkingdrum]

/assign

[nicokaiser]

Great to hear!
Isn't clipping also bad for negatives, since the uncorrected areas might turn out .. green? .. then? Anyways, clipping in negatives can (and should) be avoided anyways, so this might be far more interesting for slides.

[thetalkingdrum]

On the Negative pipline side of things: I looked a bit into it, and it seems like it could add some false color casts to the film base because of the way these algorithms work. But could be worth looking more into.

On the Slide side it was both easy and harder than expected. Easy because LibRaw already has this feature built in, but a bit messy because of the way the slide pipeline is set up. I made something that works and I will probably put up as a draft PR soon.

[thetalkingdrum]

OK put up a draft PR. If you could try it out that would be helpful, as the file you posted was the only image I had available to test on.
@nicokaiser

[nicokaiser]

I think (!) I have one or two more pictures like this, I’m happy to try out. Is there a macOS build in the PR? I can also try on Linux otherwise…

[thetalkingdrum]

You will have to build from source on either Linux or MacOS or wait until the pr becomes part of an official release. If you haven’t done it before som ai chat can walk you through it and you’ll have it up and running in no time

[nicokaiser]

I just tested with some more slides from that roll that was scanned a bit too hot. And it is much much better:

(On the last slide I exaggerated the Highlight Density a bit so the effect is more visible)

I also think the solution with optional Highlight Recovery is good, so one can decide if they want to use it or not. I could not see a difference between the algorithms though.

[thetalkingdrum]

That looks really good actually. Glad it worked. The reason to have it optional is that when the clipped pixels are not something neutral like here, but a colored light like a neon sign or similar, the clipped pixels would turn white-ish as the algorithms only work if the clipped pixels are close to neutral. That is part of the reason it would not work well with negatives as I understand it. The orange mask is a colored light.

[seanharding]

Sorry, put a couple of comments here on the issue rather than the PR. Deleted and re-posting them in the right place.