NVIDIA takes on AMD FSR with their new open source Image Scaling

Quoting: MohandevirThe FSR maintainers should have a look at DLSS and see what can be integrated that will boost FSR performances... Isn't it open source, afterall?

This is not possible for two reasons:

- The DLSS SDK is proprietary. In fact, even Quake 2 RTX (maintained by NVIDIA) doesn't include DLSS support because integrating the DLSS SDK would violate the GPL.
- DLSS requires motion vectors, which makes it much more difficult to integrate in an existing engine. In contrast, FSR is a post-processing shader that can even be injected into applications that weren't designed for it (at the cost of UI elements looking worse than with a "proper" implementation).


Last edited by Calinou on 17 November 2021 at 12:38 am UTC

Yeah they way I read this is:
Thanks to AMD once again for forcing Nvidia to do something that they said wasn't possible that they could've done at any time but chose not to. I know that's blunt, but this really highlights how important competition is. Nvidia could've done this at any time, but just chose not to, and thanks to the AMD FSR it has really forced their hand (this goes with other stuff too, such as FreeSync, high core-count CPUs, etc).

Edit: Looking at those images provided, the text looks better on the new Nvidia tech, but the wall textures appear to be "softer" and less detailed. Either way, its nice to have competition, and it looks like this will be really nice in some cases (being able to choose from either method is going to be awesome).


Last edited by holisticboy on 17 November 2021 at 1:24 am UTC

FSR still has some sharper texture handling but I suppose you can fix that issue with vkbasalt.

I would like them to put some more effort into distance detail enhancing rather then up in your face comparisons. Often these scaling systems break down when dealing with medium to long range objects/textures etc...

Fragmentation is well known in the Linux community. So we can easily deal with that and call it choice.

Nvidia could have done better, but no one is forced to use their solution.

I hope someone ports the image scaling, I have been using fsr with proton ge but its very buggy.

Quoting: FredrikI hope someone ports the image scaling, I have been using fsr with proton ge but its very buggy.

I never hit a FSR bug in Proton GE. Works like a charm if a game doesn't do its own upscaling like Control.

Quoting: chelobaka

Quoting: FredrikI hope someone ports the image scaling, I have been using fsr with proton ge but its very buggy.

I never hit a FSR bug in Proton GE. Works like a charm if a game doesn't do its own upscaling like Control.

Personnal experience... I'm using FSR and ProtonGE with Witcher 3. I get game freezes that don't happen on stable Proton, or Experimental. I mean, one complete freeze per hour of play or so.

Edit: But this said, for this particular game, I never tried ProtonGE without FSR (forgot that part), there is no incentive in doing so; stock Proton (or experimental) is more than enough. It might just be a regression in ProtonGE... Adding this to my list of tests to do...


Last edited by Mohandevir on 17 November 2021 at 4:43 pm UTC

Quoting: Calinou

Quoting: MohandevirThe FSR maintainers should have a look at DLSS and see what can be integrated that will boost FSR performances... Isn't it open source, afterall?

This is not possible for two reasons:

- The DLSS SDK is proprietary. In fact, even Quake 2 RTX (maintained by NVIDIA) doesn't include DLSS support because integrating the DLSS SDK would violate the GPL.
- DLSS requires motion vectors, which makes it much more difficult to integrate in an existing engine. In contrast, FSR is a post-processing shader that can even be injected into applications that weren't designed for it (at the cost of UI elements looking worse than with a "proper" implementation).

I meant Nvidia Image Scaling... DLSS came to my mind as a reflex, but yeah, it was the wrong appellation. Are you meaning that NIS is also locked behind proprietary stuff nonetheless?

Quoting: AnanaceWait, for that comparison image, are Nvidia saying that a lower resolution image upscaled with DLSS gives a better resulting 4k quality than a native 4k render?

They seem to run some edge-enhancement on the DLSS image and that can make some thing look better while in reality wildly distort the image. The technique was commonly used on low bitrate DVDs (and still used on some BR:s because why not) where it often resulted in ghosting), but it works excellent on things that are blocky/line-based by nature such as text which is what we see in the example image from nVidia.

Quoting: F.UltraThey seem to run some edge-enhancement on the DLSS image and that can make some thing look better while in reality wildly distort the image. The technique was commonly used on low bitrate DVDs (and still used on some BR:s because why not) where it often resulted in ghosting)

IMHO this could only be done by the player, not by anything on the DVD...?

Quoting: Eike

Quoting: F.UltraThey seem to run some edge-enhancement on the DLSS image and that can make some thing look better while in reality wildly distort the image. The technique was commonly used on low bitrate DVDs (and still used on some BR:s because why not) where it often resulted in ghosting)

IMHO this could only be done by the player, not by anything on the DVD...?

No it was done during the mastering process. I forgot to mention that what they usually did was to first apply a quite powerful filter to remove all fine details, spots and grain so that they could reduce the bitrate and then they applied edge-enhancement after to sharpen up the picture to fool your brain into thinking that there where more details on not less.

"Good" examples are the old DVD releases of Apollo 13 and Tombstone, if you have either on DVD then check them out. You will see small halos on characters left and right and if you study a still frame you will see that they have filtered out much of the details.