Understanding crosstalk matrices #662
Replies: 8 comments 20 replies
|
Private FB group, but I think I confirmed part of my suspicions: https://www.facebook.com/groups/1514948298527146/posts/7332629553425629 LS-8000 has LED peaks at 470, 525 and 650nm. I looked at the commit that added the 500T matrix #340, and it's based on the peak wavelengths of the dyes rather than a wavelength an LED might be isolating. |
|
Built in matrices (film stocks) were just approximated from film spec sheets - not measured. |
|
Not going to pretend I can answer all of this, but my general idea is the same as yours. If not, you are mixing your scanner's crosstalk/influence together with the film stock. The best way to do this is probably with a color chart as discussed in the issue thread about camera sensor calibration here a few weeks ago: #561 And you are right that your Coolscan does not need any sensor calibration matrix as it uses a monochrome sensor. But it needs density mixing imo. Usually the reds and greens are unrealistically strong and need some mixing back in to look natural. The narrow band toggle also does a decent job at this. But yes, I would treat the mixer more as a scanner calibration tool, or scanner+film stock tool if you want to get more accurate. I haven't shot a ton of Vision 3, but when I have cross-processed it in C-41 I always get some color casts / crossover that I can't seem to completely get rid of. If shooting landscapes and whatnot this is no big deal, but when shooting portraits it's easy to see that the skin colors are not quite there. That has been my experience anyway. Maybe C-41 in ECN-2 is similar? |
|
And yes, it’s a per roll thing. Not a per image kind of setting. I think the docs say you need to re-do the roll analysis after changing to another density matrix. Though you can use it without doing the roll analysis. |
|
How narrowband scanner LEDs, film dyes, and color space primaries all interact is something I've puzzled over a lot myself. I still don't have it fully figured out. It's genuinely confusing. |
|
Crowdsourcing scanner + film would definitely be doable and something I hope happens. Lightsource + camera + film is a bit harder in that there are so many different possible combinations. |
|
I figure light source + camera is super easy for users to do now (literally take r, g, b exposures of your light source I believe), so light source + stock is the only things that matters. I think. Might be underthinking things. But my interpretation is if your camera is calibrated to the light source, all cameras also calibrated to that source should give the same output and amount of crosstalk on the dyes. |
|
I think the sensor matrix is only really easy to make if you have a light source like the ScanLight that can turn on/off the individual colors. For white light it would not be that easy. But also not that needed, as the white light kind of averages out the sensor crosstalk. |
|
Right, I think it's basically impossible to do a wideband white light calibration. I imagine you would need the entire spectral distribution of the light, and to somehow isolate how much of varying wavelengths the bayer filter lets through for each color. Like you have two non-discrete, continuous distributions to mash together that need special equipment to measure. That and with a Scanlight v4 "only" being $100 seems the juice isn't worth the squeeze. I figure any people doing this level of finagling are likely fellow sickos that are willing to splash a little bit for a narrowband light. |
|
I know I would definitely go with the scanlight if I were to jump to camera scanning. 100$ honestly feels quite affordable. |
|
I had a little session with Claude working on this. Here's what I did:
name = "Kodak VISION3 500T 5219" This is the output. I feel mixed on the results. It kinda makes the greens radioactive, whereas the current built-in matrix shifts greens toward yellow in comparison. Skin tones do seem a bit more natural to me on this matrix, at least darker skin tones. To be honest, I don't know if this scaling dye densities is the way to go. Computer told me that negpy intentionally peak normalizes. So not sure if anyone has feedback, or happens to have my exact same setup of 500T in ECN-2 with a coolscan that could throw this on a color chart. TBH, this probably wasn't worth the effort and you could probably get better results faster by just dragging sliders with an image of a color chart. I don't have one though so figured I'd try this |
|
I did a different experiment, where I took one of my color chart shots, and had Claude write code to derive a crosstalk matrix based on the chart photo and the expected color values of the squares. The results of that (combined with an earlier stab at the sensor calibration feature) are in the fourth image in this comment. But I don't know of a good way to do it without a color chart, and I'm also not 100% convinced that the resulting matrix is "correct." |
|
@akesher123 have you used Claude skill that's in negpy repo for deriving the matrix. It was used for all bundled film matrices. I built this skill and then just fed it the spec sheets :) I think the effect they make depends a lot on scanning setup. I don't do camera scanning, I have plustek 8100 and more recently got Nikon Coolscan V and rescanned some photos. and I notice that those specsheet-derived matrices look better (slightly more natural) on same negatives that were scanned on Coolscan. I think calibration based on the color chart shot & your scanning setup is the only way if you are after having 100% accurate colors. But even then probably the calibrated values might be different roll-to-roll depending on film development, one roll being slightly older than other etc. even if it's same filmstock... but also I think it's not what film photography (or photography in general) is about. You should make the colors/tone look like you want them to look. It also shows with lab scans. You can get Frontier scans (so kind of the benchmark when it comes to film scanning) from multiple labs and then each one will look slightly different because it's up to whoever was operating the keyboard |
|
@akesher123 can you upload some of those 50d scans? |
|
@seanharding yeah, this was the only lab in the area that offered true ECN-2. They stopped offering it right when I dropped off, I'm guessing they just didn't feel like spinning up a run for it. @marcinz606 I'll rescan when I get home, I deleted the source scans since none of them were really archival tier images... But I totally agree on some of the color films. Gold and ColorPlus do great in ECN-2, especially in Negpy. |
|
https://drive.google.com/drive/folders/1oOO5bkcROFd50Tz8jHwKlATAMYjF_jYj?usp=sharing I uploaded some 250D scans here. Home-developed in ECN2, scanned with Plustek 8100 2.5 years ago. |
|
These look great! Both tone and colors land spot on. Not that surprising I guess, since you probably have used your own scans quite a bit when tuning defaults. Are these ECN-2 or C41 developed? Edit: Ah home developed in ECN2. |
|
I would say yours definitely convert more naturally than my 50d right now. The crosstalk matrix for 250d looks more natural than the default as well, especially with batch analysis. These inversions are quite nice. The skin tones MAY be a little red for me on both 192 and 350? Overall if these are the results I got from my 50d I would be very pleased. I'd be interested in seeing how these convert if you scan them on your LS-50. That would really help nail down the theory of LED spectra mattering between scanners for crosstalk. |
|
I have bunch of non-developed AHU 250D and 500T, waiting till i gather up 15 rolls so I can mix up the chemistry and develop them in 3 batches of 5 |
Uh oh!
There was an error while loading. Please reload this page.
Lately I've been using Negpy for respooled cinema film. Sometimes I get odd results, and sometimes crosstalk matrices can help out, sometimes they produce even odder results. I'm trying to understand them more and have a few questions.
For context: I am scanning various stocks on a Coolscan V and 8000. Now I am doing all my own development in ECN-2, regardless of shooting cinema film or standard C41 film. Negpy actually works quite well with cross-processed gold for instance.
So a few questions:
Do crosstalk matrices depend on the scanning source? I know narrowband DSLR scanning just had sensor calibration matrices added in. However, my understanding for Nikon-style scanners is they are narrowband, monochrome CCD scanners and don't need sensor calibration.
What I am confused on is: does the wavelength of the light source affect the "correctness" of the crosstalk matrix? I ask this in the context of Negpy having built in crosstalk matrices for different stocks. If these were generated based on the results of a Plustek, for instance, are they still valid for a Nikon scanner? If Nikon uses a shorter red wavelength, that would mean more green crosstalk as we get closer to the green peak?
If this is the case, does it really make sense to have these matrices bundled with Negpy? If I'm understanding it right, they still need to exist as a "scanner + stock" matrix, not just stock.
Also, are the Vision3 crosstalk matrices assuming C41 development? I've struggled to find a concrete answer here, but my understanding is C41 and ECN-2 produce different dyes in the negative, so different crosstalk matrices may be necessary?
And r.e. the Vision3 500T matrix, how does that factor into color temperature? Is the proper way to tune a matrix for 500T to shoot a color chart in broad daylight, manually adjust the color temperature first to make it daylight balanced, then tune a matrix for it?
Lastly, how are crosstalk matrices intended to be used with roll averages? I've had great success with roll average improving things, often times it seems to end up offsetting some of the odd color casts that crosstalk matrices may help? Just wondering what best practice is here.
All reactions