Nvidia Image Scaling vs DLSS

[Borgus Bohr]

Is Nvidia Image scaling (in Nvidia Settings), just as good as DLSS? I know from a technical stand point DLSS is kind of better. But really, when just playing, isn't image scaling just as good? Just as sharp? For free, and applicable to all games. What do you think?

 

I've been testing games running at 1440p, with image scaling on a 4k monitor for a while, and to be honest, image scaling is pretty great.

[Raytsou]

Image scaling is traditional scaling, also known as super sampling anti-aliasing (SSAA), which renders the frame at a higher virtual resolution, then fits it to your screen. Nothing will look better than this, but this is very computationally inefficient.

 

DLSS is AI upscaling with temporal filtering. In other words, it has an algorithm that looks at the current frame, along with the few frames leading up to it, and guesses what a higher resolution version of it might look like.  This is usually not noticeably less accurate but significantly faster.

[Raytsou]

I suppose a more satisfying answer would be to also explain a bit about anti aliasing, or the problem of representing curves within a grid. I mentioned SSAA and that it's very computationally inefficient. Most of the later methods try to achieve the same effect by trading off accuracy for speed.

 

Most popular is multisampling anti aliasing, or MSAA. Your GPU will take look at only the parts of your screen where you'll notice a difference (where there are curves) and selectively scale those parts to look better. No discernable difference at all from SSAA but is much faster.

 

We also have FXAA, which is just as popular. In this method, we have the CPU blur the image outputted by the GPU. Very inaccurate but extremely fast.

 

We next have TXAA, which instead of rendering at a higher resolution, will look at the few frames leading up to the current frame. Looked okay bit was computationally expensive. Often would be used in conjunction with MSAA.

 

We then have deep learning super sampling, or DLSS. DLSS 1.0 is pure AI image upscaling. 2.0 conceptually combined 1.0 with TXAA for best accuracy and speed.

 

[Borgus Bohr]

On 7/31/2023 at 11:36 AM, Raytsou said:

I suppose a more satisfying answer would be to also explain a bit about anti aliasing, or the problem of representing curves within a grid. I mentioned SSAA and that it's very computationally inefficient. Most of the later methods try to achieve the same effect by trading off accuracy for speed.

 

Most popular is multisampling anti aliasing, or MSAA. Your GPU will take look at only the parts of your screen where you'll notice a difference (where there are curves) and selectively scale those parts to look better. No discernable difference at all from SSAA but is much faster.

 

We also have FXAA, which is just as popular. In this method, we have the CPU blur the image outputted by the GPU. Very inaccurate but extremely fast.

 

We next have TXAA, which instead of rendering at a higher resolution, will look at the few frames leading up to the current frame. Looked okay bit was computationally expensive. Often would be used in conjunction with MSAA.

 

We then have deep learning super sampling, or DLSS. DLSS 1.0 is pure AI image upscaling. 2.0 conceptually combined 1.0 with TXAA for best accuracy and speed.

 

Amazing reply! Thanks