techspot Nvidia drivers reveal AI-powered downscaling feature called DLDSR

[UltraMega]

Quote

Dynamic Super Resolution is Nvidia’s method for letting users easily downsample games. It renders a game at a higher resolution, then shrinks it back down to the native resolution of your monitor. This results in extremely effective anti-aliasing but only performs well if a GPU has the extra horsepower to run at that higher resolution. It’s a good way to make older games with outdated anti-aliasing technology look cleaner.

 

Deep Learning Dynamic Super Resolution (DLDSR) uses RTX graphics cards’ Tensor cores to make this process more efficient. Nvidia’s announcement claims using DLDSR to play a game at 2.25x the output resolution looks as good as using DSR at 4x the resolution, but achieves the same framerate as 1x resolution.

Latest Nvidia drivers reveal AI-powered downscaling feature called DLDSR

WWW.TECHSPOT.COM

---

Dynamic Super Resolution is Nvidia’s method for letting users easily downsample games. It renders a game...

Will be interesting to see where this goes. 

 

 

[mouacyk]

11 minutes ago, UltraMega said:

This results in extremely effective anti-aliasing but only performs well if a GPU has the extra horsepower to run at that higher resolution. It’s a good way to make older games with outdated anti-aliasing technology look cleaner.

It's really sounding like:

Render at 2.25x native resolution -> AI downscale to native resolution -> display

rather than

Render at native resolution -> AI upscale to 2.25x -> downsample -> display

 

What does AI downscaling provide in the former, that normal downscaling doesn't do? Proper edge detection for meaningful anti-aliasing?
It is unlikely to be the latter, because motion vectors or game-specific training would be needed.

[UltraMega]

1 hour ago, mouacyk said:

It's really sounding like:

Render at 2.25x native resolution -> AI downscale to native resolution -> display

rather than

Render at native resolution -> AI upscale to 2.25x -> downsample -> display

 

What does AI downscaling provide in the former, that normal downscaling doesn't do? Proper edge detection for meaningful anti-aliasing?
It is unlikely to be the latter, because motion vectors or game-specific training would be needed.

As I understand it, it's basically like using DSR and DLSS at the same time and that's basically about it. DSR will give better AA results and so will DLSS so I guess this would be like super extra (probably overkill) AA. 

 

I could see this being great for 1080p and 1440p but at 4k the need for this level of AA is just not that great so the performance trade off probably wouldn't be worth it. 

 

I wonder if this is going to be a game specific feature like DLSS or if it will be at the driver level. If it's the latter, that implies that Nvidia could make DLSS a driver level implementation if they wanted to. 

[mouacyk]

DLDSR is coming Friday at the driver-level, as new options in the Control Panel and the old DSR will move to a legacy section. DLSS will never be driver-level, because it requires motion vectors to be calculated by the game engine.

[UltraMega]

I wonder what's different about this that let's it work at the driver level. 

[mouacyk]

The AI downscale is likely using a neural network trained to eliminate shimmering and edge aliasing, and thus needs the tensor cores to run. However, being a downscale process, you could probably achieve just as good a result by using spatial algorithms that do not need the tensor cores.

 

Honestly, when you're down-sampling, there won't be much difference between a source image that is 4x or 2x. Even at 2x, a simple fast sharpen filter would have been sufficient.

Edited January 12, 2022 by mouacyk