About ColorDifference/DeltaE function return meaning

[explosion-mental]

tldl; Would be good to standardize the results from 1%-100%/0.1-1.0 or just add some documentation about the result types (for complete rookies c:) for color difference traits.

Hi, I've been using the delta_e 1994 manually implemented and I've been guiding the result of such function with this simple and easy table.
However, when I switch to palette it seems that other methods, such as, improved_difference (meaning, CIEDE2000) and other functions/traits don't behave that way. Only .distance() works with the table I've linked before, and I'm not sure what to look for with other implementations.
This is very interesting as well, and If I understand correctly, it modifies the result of the standard function (from a quick glance at the doc) and I'm not sure how to handle that.

Background and use case: I use that as a configurable threshold in wallust in which I use the table mentioned to explain users how a threshold works (although is not a variable that many modify).

[Ogeon]

Hi, some sort of table like that could be a nice addition, where it's possible. They are currently implemented to be proportional to the scale of the source values, but it's not impossible that a bug or inconsistency made its way in. The CIEDE2000 implementation may be in need of a checkup, due to its age. Do you have any examples of when they differ and what you are comparing?

The traits with "improved" in their names are indeed tweaked variants of the original formulas. The idea with that research paper is to improve the average result by slightly modifying the output. You can ignore those if you prefer the original formulas.

Also, wallust looks really nice! 👀 Cool!

[Ogeon]

Hmm, maybe my phrasing wasn't the best. What I mean is that each formula changes the meaning of "20", or "8", or any other value along the scale. If you use a table for formula A, it's not necessarily right for formula B, so it will likely need to be adjusted between formulas. Ideally, a difference of 10 should look half as large as a difference of 20. If that's the case depends on how well the formula predicts the difference.

[explosion-mental]

Thats the point Im trying to understand. Im not sure how to tweak the threshold since I dont know what "scale" follows such operation, as the only ref I have is for good old deltae.

[Ogeon]

Ah, I think I get it now, sorry. Do you want to make it possible to choose both color difference formula and threshold value? While still having the same scale for the threshold for each combination? Or just to switch to a new formula, but keep the old scale?

I'm not sure that's easily possible. You can perhaps try to analyse the values and find a good enough conversion (like the one you already suggested), but some aspects of the formulas aren't directly comparable. Especially how the older formulas don't correctly take saturation into account, as the article you linked mentioned.

[Ogeon]

The table in the article is based on one person's subjective impressions, so it's not really based on a scientific process. If you run a number of color pairs through the algorithm you want to use and try to describe how different they are, you will have a similar table.

[explosion-mental]

The table in the article is based on one person's subjective impressions

Right... Well this is embarrassing, there is no "scale" then. That is up to the observer(s).

Thanks for all of the clarifications :)

I'll have to test and figure out wether to keep presenting the threshold to users to configure (meaning trying to keep it consistent) or hardcode one that has good enough results (and maybe iterate from a higher to lower threshold)

[Ogeon]

A table can probably be made, and it's probably going to look similar to that one for that formula. You may be able to find some way to have your own "meta scale", if you calibrate it to a number of differences. Like say that 100 is the difference between white and black or two other very different colors, 50 is some difference between, etc. Or take values from one formula (CIEDE2000 with or without the improvement, for example) and try to fit that curve to values from the other formulas.

I hope you find a nice solution and sorry for not really having a specific one.

[explosion-mental]

Forget all said, I've hesitated to use this lib because a knowledge gap, but the features and conveniences it gives are pretty good (although it doesn't have a complementary trait, which shouldn't be that hard to impl) sorry for any inconveniences and thanks for the crate :)

pd: apparently, you can use: type Color: Mix<Scalar = f32>, which is how I solved an issue.

[Ogeon]

Good that you found the solution! I wasn't fast enough. The color types can use different number types, so the traits need to be open for that as well.

Also, don't worry about asking questions. I'm happy to help if I can, and each question highlights a point where a knowledge gap could potentially be closed. The same for suggesting features. It makes it easier to know what to add next.