Well there's different ways to do that. A gaussian blur smooths a noisy input, but also destroys detail. There's definitely potential for smarter ways of denoising by taking various cues and using various reconstruction. It's the kind of thing ML should be good at. Seeing the result of upscaling from DLAA2, I imagine ML processing of noisy lighting data combined with simple geometry cues to be potentially very good. Maybe clean, crisp lighting of a whole scene with <10% coverage good.
I'm not sure ML gives an advantage - the problem might be too simple to require ML. If we compare a task that can already be addressed with simple low pass filter in comparison to the task of labeling hand drawn letters for example.
On the other hand the given samples can be very bad, making it difficult. In an earlier Q2 VKPT video they showed how the results would look without light sampling (next event estimation). This means rays are only shot randomly, hitting a light just by small chance.
While this would converge to the correct solution with many samples just as well, this input is not good enough to give any usable result - even the filtered image was mostly black with some flickering blotches.
Probably, ML could not do so much better here, or if so - the trained data might only work in the same scene and conditions.
It seems more important to get better samples in the first place. We want a good representation of the direct lighting if it dominates. If direct lighting is absent, the indirect light is hopefully low frequency and still variance is low enough to keep the filters working.
But there are always difficult situations where it all breaks down. Games will be designed to avoid this, and we get a new form of restrictions that no kind of denoising magic alone could solve.
Intense skylight coming through small windows would be a good example for such limitation. Simple next event estimation is inefficient because it's unaware of the small holes and where they are. So it would make sense to treat the holes similar like area lights, which would require to have some kind of portal polygon data structure with high probability of light passing through.
Another example would be street and traffic lights at night in city scene. With so many lights, we might want to combine them similar to the Lightcuts algorithm.
I guess, after the recent leap in denoising we may not get so much more out of it then what we already have. But using its variance measure as feedback to drive different sampling strategies is eventually one interesting future direction. The idea is not new, but the realtime constraint could spur some new research that was not attractive for offline.
Seeing the result of upscaling from DLAA2
It's too proprietary to draw conclusions for me. I assume most of the better results come from the addition temporal jittering so the high res information is there in a standing image. That's neither new nor machine learning just because it uses Tensor cores.
Results are good and they can call it whatever they want.
But it's not that i do not see any use of ML in general for games. Would be awesome if it could add detail to the wood in Dreams, for example.