The issue with color to grayscale conversion for human consumption is in most cases there is no well-defined ground truth. People don’t see in grayscale, so the appearance preservation approach doesn’t work. And the source image was most likely heavily color corrected to match certain aesthetic. So the problem becomes “to preserve as much information, both content and aesthetic, within constraints of the target grayscale medium”.
The bottom line is, use some standardized conversion (like described here — just to avoid surprising users) if images don’t actually matter, some contrast-preserving method if content matters, and edit creatively otherwise.
Agreed. For similar reasons a lot of B&W photographers use different color filters to achieve the look they want, rather than only take the film's native rendition how it is.
I don't do film photography but even then I often experiment with different grayscale filters to achieve different looks. I remember reading a book a long time ago that recommended different conversions for landscape photography and portrait photography with faces.
I ran into this subjectiveness in graphics recently. Thought I was doing the "correct" thing blending in linear space but turns out blending in SRGB looks a lot better for certain applications and that's what most popular applications do.
The thing that is difficult to "math," is that we perceive color in a certain way (if you ever look at the CIELAB[0] space, that's based on human eye perception). So there's a lot of "it just don't look right." involved.
I have found that getting weighted LUTs that have been extracted from some process (math, context measurements, user testing, etc.), and simply applying them in the conversion is how you execute the conversion, but generating the LUTs is the tricky part. It's not always best handled by a formula. I guess you could really go crazy, and generate the LUT on the fly, as a per-pixel conversion (we actually did something like this, for RAW conversion).
I've run into this as well. Problem is that linear RGB is most definitely not a perceptually uniform space, so blending in it frequently does something different than you want. Use linear for physically based light and mixing, but if you are modeling an operation that is based on human perception it is going to be completely wrong.
The dark irony then, is that sRGB with its gamma curve applied, models luminance better (closer to human perception) for blending than linear does. If you can afford to do the blend in a perceptually uniform space like oklab, even better of course.
It's possible I wasn't specific enough when I said "graphics". Typically I blend in CIELAB when interpolating between colors for visualizations (eg data science).
But I'm unaware of rendering engines that do alpha blending in something other than linear or SRGB. Photoshop, for instance, blends in sRGB by default, while renderers that simulate light physically will blend in linear RGB (to the best of my knowledge).
It depends on the GPU and the implementation, but I personally would not want to spend the compute on per-pixel CIELAB conversions for blending.
Since this is all perceptual anyways, go ahead and compare some color formats for yourself. I'll say that for me Oklab is the most perceptually accurate at lightness representation
The bottom line is, use some standardized conversion (like described here — just to avoid surprising users) if images don’t actually matter, some contrast-preserving method if content matters, and edit creatively otherwise.