AMD FSR antialiasing discussion

  • Thread starter Deleted member 90741
  • Start date
Nvidia hasn't stopped any game from implementing FSR thus far. In contrast to AMD who seemingly has stopped some of them from implementing DLSS.
A dev implementing something is different than it being modded in.
Also would have to go look at just released Nvidia marketed games to see if they got FSR2 as soon as it became available.

I could really see them paying and keeping FSR2 off the game at launch.

But as I said, I'm willing to wait for couple DLSS versions to see what they do.

Walt, are people really surprised at what I think marketing deals may comprise timed exclusivity? This isn't something new, and happens all the time in games.
 
A dev implementing something is different than it being modded in.
Also would have to go look at just released Nvidia marketed games to see if they got FSR2 as soon as it became available.

I could really see them paying and keeping FSR2 off the game at launch.

But as I said, I'm willing to wait for couple DLSS versions to see what they do.

Walt, are people really surprised at what I think marketing deals may comprise timed exclusivity? This isn't something new, and happens all the time in games.
They won't do anything because FSR works on their h/w. And if its adoption will lead to more games opting for more advanced ray tracing for example then it's also a win for them.

Also DLSS is still better so there will still be both demand and reason to include it along with FSR2.
 
The Interface description being open and known is very different.
Example: youtube may have api that everyone knows, doesn't mean they don't purposely break things that use it.

Nvidia has marketing with a lot of games, so I don't see them being open to people easily modding in FSR2 if they can stop it.
Having similar interface is one thing, paying for marketing is another.

I personally won't be surprised if future DLSS versions won't allow easy modding.
But we'll see, something we'll have to revisit.

Rather, Nvidia making it easy to integrate upscale technologies from multiple vendors.

 
Wouldnt see why not either, AMD aint competing with DLSS to begin with, and FSR2 runs (faster) on NV hardware.
FSR 2.0 is definitely competing against DLSS. you still seem to think "AI" makes DLSS somehow special or different - it doesn't. Only thing using neural net changes is how you pick and weight your samples and even then it only changes the reasoning behind each choice - in one it's determined by algorithm, in other by neural net
 
Tried it on 3090 and there are ghosting issues compared to DLSS. The player's car is most egregious and quite likely a bug because DLSS had similar issue in the past, even if not to the same degree. NPC cars and NPC legs leave a ghosting trail which I think is FSR's fault. Other than that there's some pixelation issue in the distance, disocclusion artifacts with melee weapons and probably the same with some vegetation. otoh FSR does not have those weird flashes in evening/nights like DLSS with RT on.

But I'm quite surprised how well the Ultra Performance mode works with FSR and seems more stable than DLSS's. Not sure if it's taking the same input resolution as DLSS but there is peformance improvement compared to the performance setting so it's upscaling from a smaller resolution.


 
FSR 2.0 is definitely competing against DLSS. you still seem to think "AI" makes DLSS somehow special or different - it doesn't. Only thing using neural net changes is how you pick and weight your samples and even then it only changes the reasoning behind each choice - in one it's determined by algorithm, in other by neural net

FSR2 is akin to TAA/TAAU. DF has explained before why fsr isnt a competitor to dlss in multiple Directs they do on mondays. FSR is complementing dlss in the pc space.
 
FSR2 is akin to TAA/TAAU. DF has explained before why fsr isnt a competitor to dlss in multiple Directs they do on mondays. FSR is complementing dlss in the pc space.
And DLSS is just another "TAAU".
Temporal, check. AntiAliasing, check. Upscaling, check. Using pretrained neural net instead of algorithm(s) in the middle doesn't make DLSS any different. Or XeSS for that matter.
You'd probably argue moon is a star if DF said it.
 
And DLSS is just another "TAAU".
Temporal, check. AntiAliasing, check. Upscaling, check. Using pretrained neural net instead of algorithm(s) in the middle doesn't make DLSS any different. Or XeSS for that matter.
You'd probably argue moon is a star if DF said it.

You already have custom solutions on consoles, all this dlss, xess, fsr etc wont see much or any use there at all.

Dlss and fsr arent here to compete with eachother, they are different solutions for us to choose from, and depending on what gpu/game one might be a better fit than the other.

If fsr was a competitor in the form of amd vs nv, then fsr running better on NV hw would be abit awkward.
It doesnt seem like either nv or amd are trying to prevent each tech from existing.
 
You already have custom solutions on consoles, all this dlss, xess, fsr etc wont see much or any use there at all.

Dlss and fsr arent here to compete with eachother, they are different solutions for us to choose from, and depending on what gpu/game one might be a better fit than the other.

If fsr was a competitor in the form of amd vs nv, then fsr running better on NV hw would be abit awkward.
It doesnt seem like either nv or amd are trying to prevent each tech from existing.

For multiplats, fsr2 seems to be the most "bang for the buck". As they can use it for pc and consoles.
 
It's the same work using default FSR 2.0 settings, but from a developers POV more work implementing FSR 2.0 additional inputs for Reactive mask and Transparency mask. I also wonder whether pc/console ports for these additional inputs would use the same settings or require different values for the best visual impact.

IIRC, I don't think TSR has these additional inputs.
 
AMD will (have to) come with their own AI/ML acceleration in the (perhaps near) future if they want to compete with Intel and Nvidia. Trying to compete with or even replace TAAU or other custom hand-coded solutions is nice, but for future projects AMD will look in the same direction as the rest does. Same story for RT.
This is Digital Foundry's conclusion.

As they can use it for pc and consoles.

Thats the thing, as was discussed in a Direct and other places, FSR2 will see limited use (if any at all) in the console space. On consoles, most developers, in special AA/AAA space have their own hand-tuned solutions, that often perform better than FSR2 does. Also, FSR2 does eat into compute performance, seemingly more so than other solutions, which is a consideration when your dealing with low-end GPU solutions. The min requirements for FSR2 on pc are quite 'high'.
 
Yeah but it's essentially the same work for DLSS so if you add one there's really nothing stopping you from adding the other - besides contractual obligations.

Yap, best "bang for the buck" is to make the engine temporal upscaling friendly. From here the implementation is one hour or so for DLSS ( and XeSS?).
 
Using pretrained neural net instead of algorithm(s) in the middle doesn't make DLSS any different
I wouldn't compare TAAU with the second gen reconstruction technologies, such as TSR, FSR 2.0, DLSS 2.0+ and XeSS at all.
The second gen technics are built around minimizing losses by getting rid of the lossy color clipping heuristic, optimizing resampling, preserving and resolving subpixel details.
FSR still relies upon color clipping, so some of TAAU shortcomings are still here. The pixel locking heuristic is to resolve thin details, which are otherwise removed by color clipping, but it relies upon the depth disocclusion mask and can introduce ghosting artifacts on the inner geometry parts. Judging by DF's analysis in GOW, pixels can be locked on parts that are not affected by the disocclusion mask (such as Kratos hands or on water), so there is nothing that can prevent ghosting from happenning on these parts expept for developer's provided masks. Overall that looks like a bunch of fragile hacks to me, these hacks can fix certain things, but break others, so games have to be carefully tuned on a case by case basis. Instead of building 1000 heuristics to handle different scenes in games, you can train a neural net. Understanding what is going on in frame, what parts of image to blend and what parts are better not touch is a key to quality and ease of integration and neural nets have proven to be the right tool for the job (vs handcrafted expert-based algorithms).
 
Back
Top