DLSS literally has a super computer to train models to create accurate AI models to do upscaling for games. you are horribly mistaken https://www.nvidia.com/en-us/geforce/technologies/dlss/
That page is exactly the marketing speak I'm talking about, too. Watch the technical explanation instead.
It's not upscaling. The rendering is still an alternating 2x2 grid.
2x2 deinterlacing is TAA. DLSS goes beyond that, and 2.0 really delivers to the point it looks better than TAA at native resolution.
DLSS is TAAU, it's still temporal, based on a 2x2 grid. 2.0 is simply a ML implementation. It doesn't upscale single frames, it doesn't add additional detail to textures. It's "temporal upscaling", which is pretty much just a fancy way of saying deinterlacing.
If discount the entirety of the implementation responsible for the quality improvement as "simply ML", then it is just fancy a deinterlacing indeed. Just like every ML breakthrough is just a fancy matrix multiplication.
But the fact remains that DLSS 2 is friggin amazing: achieving combination of visual quality and performance that was not possible before it, and in some cases it looks even better than rendering at the native resolution. It's solving a hard problem, and does it very very well.
Yup, it is an impressive tech.
My point is rather that most of the marketing speak makes it sound like an _impossible_ upscaling tech, rather than an impressive temporal super sampling tech.
I do wonder about the "better than rendering at native", though. In true TV-set marketing fashion, DLSS 2.0 comes topped up with a free sharpening filter to make it look slightly sharper than native renders, by default.
https://youtu.be/d5knHzv0IQE?t=1240
It's still alternately rendering a 2x2 grid.
(In a similar marketing trend, Nvidia's "Deep Learning Super Sampling" is more like 2x2 grid deinterlacing, not magical AI upscaling.)