Notiusweb wrote:Hey Otoy - if you are in fact tweaking the Denoising, and it causes some differences between 2018 and 2019, maybe this reveals a greater opportunity.
Understand that you have it as an all-in-one solution, perhaps there will be use-case limitations for certain situations where it looks worse in 2019 vs 2018, and then other situations where it looks better in 2019 vs 2018.
So - maybe introduce, like the Old/New/Nishita Daylight models, variant models of Denoising to suit different rendering scenarios (ie different specular, transmission, lighting scenarios)
"AI" in itself will ultimately function on an if-then command subset at the most basic level, and it, in fact, can "get it wrong" in trying to do so.
In Biology, for example, "Instinct", which is a rehearsed biological command subset for given situations, will yes reveal errors/glitches under certain situations.
Thus, we see different species exploiting or failing from varying success rates on the instinct command subset they possess.
Denoising Blend is a great tool, it is awesome. But it is only working with one model of denoising. So, maybe give us a couple different models to choose from, will make the product more versatile.
I can try to give the option to select between 2018 and 2019 models. I'll add it in the todo list.