Sigmoid and agx #873
Replies: 13 comments 58 replies
|
Sigmoid is a no-go. AgX might get merged into filmic, but it's not going to be any time soon. |
|
What does agx do specifically that filmic does not? |
|
It has a different strategy of handling chromaticity consistency from input to output, using RGB ratios (which was actually what an old color-science version of filmic did). Plus it uses a special colorspace. But it's just another variation on the same theme, that doesn't justify another (user-facing) module. |
|
For sure. All I was able to find was that it was tone/gamut mapping that desaturates towards white so I was wondering how it was actually different |
|
that maybe the theory, in practice i can do things with agx that filmic can not do. |
|
Can you please direct me to a specific video that illustrates why you prefer it over Flimic ? |
|
So after some digging, it seems like this is the main difference: a hue preservation adjustment (a slider that goes from no hue preservation to 100%), and more desaturation in the highlights. 
Agx also has pre and post tone mapping primary adjustments, which seems overkill and out of scope for a tone mapper. The curve adjustments are different, but I doubt they allow for drastically different results than filmic. The question to me, aside from slight differences in chromaticity preservation strategies, is whether controlling the amount of hue preservation is desirable/belongs in a tone mapping module. My feeling is that you'd be against it @aurelienpierre, but I may be wrong. Idk how predictable those shifts are in practice. |
|
We need to be accurate on the terms here:
What we need to preserve throughout the tonemapping is the perceptually-constant one. The one filmic uses is the angle one, because that's the only one we know how to compute. Now if we get a proper way to compute the perceptual one, why not ? But I don't know of any way that sustains HDR signals, and I haven't reviewed what AgX does in depth. My gripe with DT's AgX is they fucked up, once again, the feature splitting by squeezing all the color controls they could in there, where they don't belong. Unfucking what tonemapping does to color, yes. Making it the one-pit-stop tool for desaturated highlights, no. |
|
I was talking about the angle, which is what we do, and as we discussed a few times already is what we know how to compute. Both states, no hue (angle) preservation and 100% hue (angle) preservation, should be known. Is it possible/desirable at all to be able to interpolate between the two? It shouldn't be an accurate Abney correction, but perhaps robust enough? I am thinking about let's say 80% hue (angle) preservation. Robust may not be the right word.. but you get the idea. Regarding smoothly fading colors towards white, which is what AgX seems to do well, filmic does fairly well at "highlights saturation mix" 0%. When you start moving towards -50%, things are quite uneven (using troy sobotka's gradient), particularly if you increase contrast (in filmic). Red and blue primaries take quite a bit longer to fade to white. Filmic v5 is much better behaved in that regard. |
|
You can't do % on angles. Angles have bitten my ass more than once. I'm too old to handwaive that kind of things, especially given that we already have reliable tools for that task, so it's a situation where you need to methodically list the flaws of the previous solution, do a serious litterature review, and decide if the new cool toy is more than just a different flavour of shit, before pusing everybody on this new path. And right now my feeling is that we can already achieve what AgX does but through 2 different modules (color balance + filmic). |
|
Seems reasonable. How about the desaturation towards white? Any reason not to use filmic V5 Max RGB? That seems the smoothest of all settings EDIT. Hue preservation |
|
So filmic desaturates towards white very gently via the reconstruct tab. Raise the transition slider, move the balance sliders to the left, particularly the grey/colorful details one, then lower threshold to taste. V7+reconstruct tab is more than enough for me. |
|
I'm thinking Filmic is doing an excellent job. |
|
How do you create these gradients? |
|
the before or after ? |
|
The base gradient |
|
https://github.com/sobotka/Testing_Imagery/blob/main/Sweep_sRGB_Linear_Half_Zip.exr |
|
La visualisation des ces 3 images me suggère que le curseur AGX-contraste à lui seul agit comme l'APN, à savoir augmention simultanée du contraste et de la saturation. AGX Curseur contraste poussée de 3 à 5.jpg Boitier APN.jpg Filmique par défaut.jpg Compare and contrast_ Filmic RGB vs AgX - Software _ darktable - discuss.pixls.us.pdf |
|
Oui mais c'est pas comme ça que Ansel est pensé : le traitement de la couleur se fait dans des modules dédiés (calibration couleur, balance couleur) et n'a rien à faire dans un mappage de tonalités. La promesse, c'était un contrôle différencié et indépendant de la luminosité et de la chromaticité. Si on commence à violer le critère de conception fondamental pour émuler un JPG, je vois pas l'intérêt de de faire chier avec du RAW déjà de base. |
|
@DominiqueFrt the result you show with agx I find very pleasing. However I do agree with AP that controls should be independent, and thus, predictable. Tweaking contrast should not affect saturation or hue, for example. In terms of highlights desaturation, filmic already does this very smoothly via the reconstruct tab. The agx primary controls (pre and post tone mapping) seem overkill for a tone mapping module. Ansel handles those in separate modules. However the hue shifts agx introduces are indeed pleasing. AP seems wary of them, and perhaps mathematically it doesn't make sense, but from a user's pov I don't see the harm in a hue preservation control. Meaning of course, 100% hue preservation by default. Other than that, it's pretty much potatoe potato, with different controls. 
|
|
Oopsie… |
|
It's not merged yet |
|
Ah ok. Whenever it is. Cool stuff |
|
Color preservation means chroma preservation? |
|
color preservation = chroma + hue |
|
So this is different than the current slider, then👀 |
|
I'm getting a 404 on the image above |
|
Ok, pushed.
Some findings (TL;DR of the dev docs):
So, in Ansel:
That's 475 more lines of codes into an existing tone-mapper instead of a whole new module of 2843 lines. Overall, I think the crowd just overhyped itself with that new toy, then some youtubers did a couple of good tutorials, and that's it. The whole "primaries" thing is nothing but a channel mixer, that we already have. The rest is very close to filmic without the chroma/color preservation mode. And for the bleached look, color balance already had everything you needed. Results:
So AgX-like default has a more even hue-wise effect than v6 without chrominance preservation (the blue row gets in line with the others), but that comes at the price of desaturating everything a lot more. |
The default settings ("safe" preset for toe & shoulder) match the CIECAM16 J curve, and the 1.18 contrast/slope sits between the ~1.16 of CIECAM16 J and the ~1.2 of Bartleson-Breneman contrast adaptation. What happens is we have grown accustomed to the previous splines that allowed to recover much more highlights in a way that doesn't really match vision. And are still available. But I'm weary of black magick that looks good and gets later replaced by something that looks better to someone else, so defaults are grounded into something objective and derivable by numerical fitting.
Because once you have put both latitudes to ~0%, you will find that you will lack something to make contrast more gentle around black and white, and that's what the hardness setting is for, pending the real sigmoid power. |
|
This is the default v8 v4 spline with white and black relative exposure adjusted. Contrast has a value of 1. This is what happens with minimum latitude values.quite contrasty, and it looks like the shoulder is much sharper than the toe: 
And this is what I ended up doing, lowish global contrast and high shadows latitude value: 
Global contrast is now very sensitive to changes. I do like the very gentle highlight rolloff. It likely does not match vision, but neither does film, does it? My point being, increasing the contrast+highlights slider i can always make the transition harsher, but with the new spline, it is tricky to make it gentle. Imitating v4 with v3, so harsher: 
And a gentler version with v3: 
|
|
Yes but remember you are looking at a synthetic image having 25 EV of dynamic range, and in practice, you want to adjust the global contrast so you get enough local contrast in the central 4-5 EV around middle grey. So, if that was a real image with a subject, that gentle global contrast would have flattened it completely. |
|
That's true. My camera does not record that Dr, of course. Perhaps it's a matter of choosing one over the other depending on the scene. However I find the V3 controls (always set to safe) more flexible. Maybe I'm just more used to them. |
|
Just tried your new perceptual options for shoulder and toe. They work like a charm. You did it again. I really like this V8 version |
|
I'm about to try the new filmic. After reading your doc:
The opposite may also be true. For the infamous sunsets, pictures of fire, etc. You may want less hue recovery AND less bleaching. In most cases I agree a coupled control is probably what you want. Something to investigate. |
|
For now, it was the only way to include AgX without changing filmic's parameters, aka database histories/presets/styles, and ensuring compatibility. Anything more will need additional params, and to bump version, from which there is no turning back. |
|
Yes, yes I saw. Let's test this first and see if there's a real need. I was merely pointing out that there's an opposite use case to blue concert lights. |
|
So I've been playing with it a bit. Looks great. The largest change is going from V7 to V8. The color preservation is quite subtle, and both ends seem good. I like that it doesn't bake in a grade. However I think decoupling hue shifts from highlights bleaching is something to maybe explore. Because of the bleaching, shifts are not very apparent. All in all great improvement. I still don't know how I feel about the new spline behavior, since I was so used to the old one, but I'll play with it some more |
|
I pushed a change that will void previous edits. The problem was the pure AgX (no color preservation) was bleaching valid (non-compressed) colors far too much, so I modified the forward and backward matrices (primaries) to compensate for that alone, even without color preservation on. The matrices have been numerically fitted to minimize hue and chroma deviation across the tone-mapping. It was not the goal, but that actually brings back a larger hue shift toward yellow, which was the initial design criterion of AgX and a request for sunsets. The numerical fit has been done in priority on skin tones and reflective colors (well within sRGB gamut). The updated color preservation removes saturation recovery entirely, as it produced non-smooth gradients in highlights.
Notice how the vectorscope gets back in line as we increase preservation. Sunset: (from https://discuss.pixls.us/t/sunset-scene-with-darktable-rawtherapee/23139)
Portrait: (from my personal collection)
|
|
@aurelienpierre : I have been using the latest and I believe you nailed it ! |
|
I have more ^^ |
|
More what? 👀 |
|
More goodies |
|
What a teaser. I'll give you one back. It's off topic, but since you're into music theory, check out the masterclasses by jazz pianist Barry Harris, and the music theory of bela Bartok's music by ernö lendvai. Incredibly cool that two people coming from radically different cultures, musical traditions and countries come up with pretty much the same theoretical framework independently. |
|
Ok, new big change pushed, I have a lot of things to say. Recap of previous episodes
Studying AgX methodicallyThe contribution of AgX is to do tone-mapping on individual RGB channels, but in a special color space, where we go by applying a color matrix, and from where we go back by applying another color matrix. The color matrix is the math result of a GUI tool called channel mixer, allowing to define the 9 coefficients of the 3×3 matrix, and that channel mixer can be re-interfaced as a primaries GUI, allowing to define the 9 coefficients from 3 scaling and 3 rotations. Whether we grab the the 9 coefficients from an ICC color profile, from a channel mixer GUI or from a primaries GUI, when that reaches the pixel, it's the same product between a 3×3 matrix and an RGB (3D) pixel. The question becomes: how to build the color space for better channel-wise tone-mapping ? Aka how do we choose the coefficients for the input and output matrices ? My first idea: create a synthetic set of colors, record their initial Ych coordinates, record Ych again after the tone-mapping, and optimize the input matrix that will minimize the hue shift. The output matrix is then the exact inverse of the input. As the first results published here show: that desaturates way too much, even valid colors. In practice, that means non-white (non-Caucasian) people get white skin nevertheless. Kodak went down that path already, never again. So, why not recovering output chroma from input chroma in Ych, same as we do for hue ? Well, because copy-pasting hue is free, but doing the same with chroma/saturation creates non-smooth gradients in highlights, and that's really ugly. So the AgX matrices need to recover saturation for themselves. My second idea: same as before, but in addition of optimizing the input matrix for minimal hue shift, optimize the output matrix too (so it's not the inverse of the input anymore), and optimize for minimal hue shift AND minimal desaturation. Saturation is more critical for us here, because that's the parameter we can't recover directly from input RGB. That's better (previous results), and also restores the yellow drift that people want in their sunsets (whose intensity can be driven by the color preservation parameter), but portraits are a bit weird even when restoring hue: keeping the saturation constant hits Helmholtz-Kohlraush effect and skin, registering near red, appears too bright. Here goes the one-size-fits-all matrix fit. Here, Darktable says "let's give manual control over the input/output matrices to users", which I saw only as a shitty UI giving 12 extra user parameters meant to please the geeks and overwhelm the others, but it actually goes beyond that. See, I had to write 1000 lines of Python for a solver that finds the best matrices minimizing desaturation and hue drift, while not pushing valid Rec2020 colors out of gamut (aka mostly creating negative RGB), and manually ensured that the results were numerically stable (invertible without being overly sensitive to noise). But I'm the color & math guy, it's no problem. Now, giving this responsibilities to the average photographer is only handing them a machine gun to shoot themselves in the foot, because they get 12 parameters that will have dramatic effect on things they don't even know of, let alone understand. (Remember: Blender AgX is a collection of LUT pre-baked by someone who knows). My third idea: given that we can't avoid bleaching, at least control the amount of bleaching we allow, so re-optimize the matrices for minimal hue drift under a pre-determined average desaturation budget. That average is computed on skin tones and reflective colors, aka colors that we know should fit in any RGB space regardless of dynamic range. And that gives us 4 sets of matrices:
Final metrics:
So you get 4 variants of AgX built in Ansel (no/low/high bleach), they are all fitted by numerical optimization under clear constraints, and the Python script to refit them with different priorities or data is on the repository, so the results can be reproduced. ResultsAll results shown with color preservation set at -100% (aka no hue reconstruction at all, pure AgX). Synthetic
Note how the vectorscope gets better radially-aligned as we increase bleaching but also shrinks (chroma scale is identical between screenshots) Portrait
Without bleaching, the lipstick becomes fluorescent. High bleaching also increases the local contrast on the jacket and on the face (reds), but we start loosing the skin saturation too much. Low or medium bleachings are excellent compromises. Sunset
As expected, high bleach restores the radiometrically-correct red in the sunset while washing colors a bit too much for my taste. Pick your poison. ConclusionForum democracy is when the Open-Source world deludes itself into thinking that more clueless opinions amounts to collective intelligence if you add enough of them. There is really nothing to salvage in Darktable. |
|
And, on these same images, what does the color preservation at 0% and 100% for each of the three different color science presets? Do they constrain the bleaching and color washing? |
|
Great idea. An elegant solution to the hue/sat decoupling. No bleach looks great for sunsets, low/mid for everything else it seems. |
As I said, it's not possible to recover from bleaching from one single parameter without destroying gradients in highlights. The color preservation recovers only hues, which means it fights back the yellow shift from the no bleach approach, and becomes barely noticeable for high bleach. But the bleaching itself is built into the 3D AgX matrices, that are solved for graduated hue vs. saturation constancy trade-off, and hardcoded as discrete cases. |
|
One more thing @aurelienpierre . People have been playing with sunsets and agx in the chat. This is a crop from @Jiyone 's edit. He says this not so smooth area around the sun is the result of filmic gamut mapping (soft clipping) to sRGB. 
He then shared the same image, but in Adobe RGB. 
The second one looks clearly better. Given that my screen only displays sRGB, I'm assuming that at some stage (element app, browser, etc.) there's some perceptual intent going on. Wouldn't something like that make sense for filmic, particularly for exporting to smaller gamut sizes? At the expense of overall saturation, but in favor of smoother gradients. 
|
|
Great work and write-up. Thoughts:
|
|
Tone mapping and gamut mapping are different thing conceptually, but in fact they are the two faces of the same coin… And the difference may not make sense at all, since they are all about remapping for display, whatever display is, knowing that it's going to be narrower than the master. Perceptual intent is AgX. In ICC profiles, the perceptual intent is actually a vendor LUT that does undisclosed black magic (and since that LUT is rarely provided into profiles, almost all profiles silently use relative colorimetric — check it for yourself, no visible difference in Ansel). Perceptual intents are possible only when you know the input and output colorspaces (like Adobe RGB to sRGB). You can't bake perceptual intents for whatever RGB to whatever RGB (being decided at runtime, and connecting through a PCS). The extra/legacy gamut mapping with chroma clamping kicks in only because AgX failed to bleach enough for display. For the sake of smooth gradients preservation, there is nothing I can do better than AgX… Several strategies were tried over the years, and, like it or not, but bleaching is the least weird of them all. |
|
I see. This was not a complaint, btw. I really like the new agx. How about manual gamut compression a la color calibration? Less desirable than the more bleached presets? Not significantly better/different? |
|
Color calibration does a vibrance-like compression of the whole gamut space, in or out of gamut, hence why it's smooth. It's meant mostly for the blue-LED case and it's kind of a desperate attempt. And here again, global gamut compression is also what AgX does, but in RGB, using special color matrices. To be accurate, the bleaching is a byproduct of the curve shoulder, not of the channel mixing we apply before and after. Those are actually a way of recovering some saturation. |
|
I'll let it (and you) rest. I do not dislike the bleaching one bit btw. |
Uh oh!
There was an error while loading. Please reload this page.
I love the redesign of darktable to ansel. It looks much better and has much better performance.
One thing i am missing and that might stop me from using Ansel is the fact that it has no Sigmoid and/or AGX.
I use those 2 all the time in darktable.
All reactions