Frenetic Pony
Veteran
It's a bit odd that they don't know which code parts to concentrate on. FXAA has been fantastically successful, and has been released in who knows how many games.
A hyper compact machine learning version might be similarly successful. Teach it to interpolate local gradients over some arbitrary neighborhood so upsampling can look better nigh as a plugin.
Wouldn't compete with DLSS but would be universal. They can compete with DLSS in other, somewhat less universal ways. Any TAA should involve history rejection and history rectification (reshading). Since both end up being guesses anyway, and bad to good implementations are often the difference between bad and good TAA, this seems a good target for AMD and a hopefully compact nn. Other than a code example of TAA upsampling, which like with DLSS needs direct developer involvement to implement anyway, that's the most logical way forward I can see for them.
A hyper compact machine learning version might be similarly successful. Teach it to interpolate local gradients over some arbitrary neighborhood so upsampling can look better nigh as a plugin.
Wouldn't compete with DLSS but would be universal. They can compete with DLSS in other, somewhat less universal ways. Any TAA should involve history rejection and history rectification (reshading). Since both end up being guesses anyway, and bad to good implementations are often the difference between bad and good TAA, this seems a good target for AMD and a hopefully compact nn. Other than a code example of TAA upsampling, which like with DLSS needs direct developer involvement to implement anyway, that's the most logical way forward I can see for them.