Do you know the hardest part about Tonemapping in games?
It is going from “most things are fine” to “but fix the overexposed pixels”.
You start with 1 simple problem, then slowly unravel a dozen of interconnected problems.
I think there is truly a market for tools that can determine a tonemap profile given the original picture. (And I can tell you histogram/color models aren’t enough.)
The hardest part is about preserving the perceived image, and they are very difficult to tune by hand.
I have worked with a lot of talented environment artists, level artists and tech artists for the past 5 years.
So far, no one I knew (me included) could come in, add or change a tonemap profile, show it to art directors/game directors, and pass in their first attempt.
There are always something off about it.
Of course there are, we are playing with numbers and curves.
The problem had always been: subjectively, what could you change to mitigate a problem, leaving the rest still acceptable.
You know, as I read your first two posts I was literally thinking "here's something ML could actually be used for and be good at". Hardest part would be figuring out how to make a loss function, otherwise it's manually reviewing each round of training and that's... Not fun.
I don't know if there is or not but I assume there's some equation for calculating how over exposed an image is. However "close to original perception" is tricky, that's the kind of thing a research team could fight with for years. Maybe calculating distance in Lch space? OKLch is all the rage lately and seems really nice. It's a hard problem but definitely doable, I think.
@tacitus not sure how to do it but if we can assign “aesthetic scores” to random pictures, we can assign proximity to image somehow (badly but probably good enough, don’t need it to solve everything automatically like Gen AI).
Add comment