Nvidia DLSS 1 and 2 antialiasing discussion *spawn*

[DegustatoR]

Since TSR will run on basically anything, could it run ontop of DLSS?

No. DLSS *is* TSR, just more advanced because of ML reconstruction.

 

[cho]

someone told me NVIDIS has post the DLSS 2 SDK to github for who signed NDA wih NVIDIA.
https://github.com/NVIDIAGameWorks/dlss_private

 

[Dictator]

There is definitely less high contrast trailing in 2.2 vs. earlier iterations - should be making a video on this after I finish my current project (so next week I think).

 

[Rootax]

So... Do we have to upgrade something, or each dev have to patch their games ?

 

[OlegSH]

Well it does raise the question of whether the problem wasn't actually motion vectors but a bug in DLSS that is now fixed.

Honestly, particles without motion vectors is a thing that has to be fixed on render engine side because there are way more solutions to this problem that can be done by renderer's means (masks, separate passes, motion vectors, etc) rather than trying to fix the ill-posed problem in post-processing.
TAA, for example, fixes the issue by simply removing too much of details across the whole screen with color clamping, that's a pretty poor solution for image detalization, but it works well for ghosting removal on particles, etc.
DLSS, TSR are the opposite, they are trying to preserve as much details as possible with neural nets or more complex heuristics, but these heuristics can break hence they need additional implementation effort from devs.

 

[OlegSH]

Do we have to upgrade something, or each dev have to patch their games ?

Devs have to patch their games

 

[Deleted member 2197]

So... Do we have to upgrade something, or each dev have to patch their games ?

Not exactly sure about the procedure, but some people claim reduced artifacts in other games through swapping a DLSS 2.2 .DLL file. So sounds like combo driver and game patch update.
Cyberpunk 2077 DLSS 2.1 vs 2.2 : nvidia (reddit.com)

 

[DegustatoR]

Not exactly sure about the procedure, but some people claim reduced artifacts in other games through a "hack" process. So sounds like combo driver and game patch update.
Cyberpunk 2077 DLSS 2.1 vs 2.2 : nvidia (reddit.com)

Driver is mostly irrelevant. You can swap the NGX DLSS DLL in the game's folder to a newer one and it may work. But this is basically the same "patching" which the devs should do.

 

[Deleted member 2197]

Curious some people notice ultra performance less pixelated after substituting the file ...

 

[Scott_Arm]

Not exactly sure about the procedure, but some people claim reduced artifacts in other games through swapping a DLSS 2.2 .DLL file. So sounds like combo driver and game patch update.
Cyberpunk 2077 DLSS 2.1 vs 2.2 : nvidia (reddit.com)

Ghosting isn't gone, but it is a lot more subtle. I am watching it downscaled to 1080p on top of youtube compression, so that it's probably not as obvious as native.

 

[dorf]

DLSS plugin for UE5 is out. Developer account needed to download but it's just a simple registration process.

https://developer.nvidia.com/dlss/unreal-engine-plugin

 

[troyan]

And the UE4 plugin, source code and documentation is downloadable without an account.

 

[DegustatoR]

 

[techuse]

At lower percentages DLSS looks slightly better. At higher percentages TSR seems to look very slightly better.

 

[Scott_Arm]

At lower percentages DLSS looks slightly better. At higher percentages TSR seems to look very slightly better.

i forget what the internal resolution % of dlss quality or balanced is. Tsr 100% is native resolution.

 

[pjbliverpool]

i forget what the internal resolution % of dlss quality or balanced is. Tsr 100% is native resolution.

Assuming the percent given is the percent along either the horizontal or vertical axis rather than the total pixels then:

DLSS Quality = TSR 66% = 1440p
DLSS Balanced = TSR 58% = 1253p (not sure this is what DLSS Balanced uses but I'm assuming it to be the case)
DLSS Quality = TSR 50% = 1080p

I took screen shots of the stills for direct comparison and a Performance and Balanced DLSS has a clear advantage while at Quality DLSS seems to have a very slight advantage. I noted that closer to the camera TSR does relatively better to the point where in Quality/66% TSR looks better very close to the camera but DLSS still looks better for the majority of the screen.

 

[Dictator]

Imo stills are not going to be the way to judge this, rather motion. Temporally accumulating to make a high res Image with no movement is old hat, even a basic checkerboard with no rejection over two frames does that. What happens when the screen moves is most important.

 

[Dampf]

Also we need to test lower resolutions as well. Comparing at 4K resolution without zooming in is pretty contraproductive since most people have a 1440p or 1080p monitor, so differences can be hard to spot.

 

[CarstenS]

Imo stills are not going to be the way to judge this, rather motion. Temporally accumulating to make a high res Image with no movement is old hat, even a basic checkerboard with no rejection over two frames does that. What happens when the screen moves is most important.

Stills are still important to judge the detail retention/reconstruction capabilities. Motion for temporal accumulation, ghosting and other artifacts.

edit: / instead of 7

 

[nAo]

IMHO the "golden scarf" and the fire particles are the elements that differ the most between the two techniques.