Right. So, what you're saying is not that DLSS is better than native, but better than native + TAA.
we live in a world where engines like UE4/5, Northlight and Snowdrop exist. these modern engines are entirely designed around using TAA. try playing a modern UE5 title on PC, and mod it to remove the TAA and just look at the raw output. what you will see is an image that has dithering artifacts everywhere, and where shadows, hair, foliage and other elements on screen are flickering and shimmering like crazy, even while standing completely still.
so, in 90% of modern AAA games, TAA is not an option, it is the bare minimum setting.
My point still stands, both ate detrimental to fine motion detail, despite being used to improve resolution. Also, DLSS isn't the same as TAA as both take information at different instances, with DLSS using ML to fill the gaps were it isn't sampling from frames. If TAA implementation in some games is worse than DLSS solution, that doesn't make DLSS as magical as some here like ro paint.
DLSS doesn't simply fill gaps by using ML. the initial steps to gather additional image information is the same, both for TAA and DLSS.
jittering of the pixel grid, accumulation over several frames, reacting to motion vectors to minimise ghosting and so on and so forth...
the difference is that DLSS is not only jittering the pixel grid, but also alternates which parts of the image it samples.
which is VERY similar to checkerboard rendering... just more extreme depending on the scaling factor you use.
the ML model comes into play when DLSS constructs the final image from the checkerboard-esque sampled frames, where the ML algorithm has learned what looks wrong in an image, and tries to change the color of pixels that it thinks look wrong. so it predicts how a high resolution version of that frame should look like, and then adjusts elements of the actually rendered/accumulated frame. it doesn't fill gaps through the use of ML, it looks which parts of previously sampled frames match its prediction the best.
all of this is of course not perfect but it is what makes it cleaner looking than FSR2, which works basically identically to DLSS, just minus the AI knowing if something looks wrong, which is why everything is always fizzling in motion with FSR2
Also I'd like to say, again, that I'm not against the use of DLSS, nor I advocate against it. On the contrary, i think that it is a very clever solution, mainly for folks with very high resolution displays with poor upscaling capabilities. I just point out and that it is objectively impossible for it to produce a quality better than native, given how the technology works. Want a better experience at higher resolution? Get more rendering power, simple as that. Native is always the top quality that a GPU can produce, there's no such thing as better than native.
well, like I said, native in modern AAA games almost always means Native + TAA, and it absolutely can beat that.
sure, if you look at something like Valorant, a forward rendered, simple looking game that doesn't rely on TAA... that would indeed be impossible to match for DLSS. but in any game with TAA, at least at 4K DLSS Quality mode, you will match or exceed the image quality in 90% of cases. at 1440p it's less but still many, probably most games.
