It is one of those technologies that is very well hidden in effect behind Youtube's compression. Given how aggressive compression is on Youtube the gap in what it looks like in front of you verses on the tube has widened quite a bit. DLSS seems to be indistinguishable on Youtube but in front of you on a monitor it clearly introduces some odd artifacts and is not native looking. The gain in performance is usually worth it but its definitely not just magic free performance with no impact on image quality.

I can see the visual downgrade.

IMO, VRS (variable rate shading), which is in all major GPU providers (NVidia, AMD, and even Intel iGPUs / Xe), provides the "upscaling-ish" performance gains that we want for the next generation.

Its much harder to see the difference between native and VRS.

https://devblogs.microsoft.com/directx/variable-rate-shading...

-----

Its not a "competitive" feature, because everyone is doing it. But it basically accomplishes the same thing: carefully upscaling results (2x2 instead of 1x1) in locations where the gamer probably won't notice.

For me personally, in Control, 720p upscaled with DLSS to 1080p looks better than native 1080p presentation. I say that with a hand on my heart, the DLSS image is cleaner, the lines are sharper, and the texture filtering is just vastly superior. I have no idea how they are doing it, because it seems to be pure black magic, the traditional logic of "upscaled can never look as good as native" has been completely thrown out of the window. It doesn't look as good - it looks better in my opinion.

DLSS is trained with a much higher resolution version of the game. So when it 'upscales' 720p to 1080p its not just using the 720p data as input but also the knowledge of how the game should look at an extremely high resolution. Basically it has a lot more (hidden) knowledge than the current low resolution frame. Hence why its more effective than a traditional upscaler.

The magic essentially works because they're feeding lots of hidden information in for the upscale - the 1080p is being constructed from a bunch of 720p frames + depth/motion vectors, etc. So it's natural that it is able to look better, and they wisely tap into the (mostly/entirely unused in games) tensor units on the GPU to do it without eating too much into your shader performance.

I'm sorry but I don't understand how is natural that could make it look better than the native version.

Care to explain more or have some good resources where I can learn about it by myself?

Well of course you'll notice, but do you notice it more than if you need to drop off to non-native resolution because your card keeps dropping under screen FPS?

Because that's where the difference gets important. If I compare to my PS4 Pro, many games "look" better because they drive the UI at full native resolution while using checkerboard rendering to keep the framerate up. The same game on my 970GTX will chug, drop frames or have all the text and UI look blurry because I need to drop the resolution.

If DLSS 2.0 fixes this issue, it's a massive improvement.

DLSS 2.0 is "just" an upscaler. There's a ton of techniques to upscaling, and a lot are way better than a shitty bilinear filter. Temporal upscaling, for example, is also a thing, and is very good ( https://youtu.be/A2A-rhCQCuY?t=605 ). Upscaling in general has been a huge area of technique improvements & innovations thanks to the xbox one & PS4 struggling so hard with 4k desires.

DLSS 2.0 is a great addition to that field, but it's far from being the only good option here. AMD has also done some things in this area, too, like Radeon Boost which was the idea of dynamically reducing resolution in response to player inputs, with the idea being if you're sitting still-ish you're OK sacrificing FPS for better quality, but as soon as you start whipping about you want the FPS more. Game engines also do similar things, and combined with temporal upscaling & antialiasing are quite compelling.

What I had been wondering about was "what if you place DLSS like technology into the screen itself?

DLSS 2.0 requires motion vectors as an input to re-project the previous frame, so an equivalent of that at least can't be implemented in the display.

DLSS is also a rather heavy computation, taking 1-2 ms per frame on high end GPUs. That's a serious amount of processing power to put into a display and doing so makes it exclusive for that use. On the GPU those tensor cores can be used for other tasks when they're not doing DLSS.

You need to add Nvidias supercomputer to the screen too then.