AMD FSR 3.1 Announced With Vulkan Support

[phoronix]

Phoronix: AMD FSR 3.1 Announced With Vulkan Support

AMD used the Game Developers Conference (GDC 2024) this week to announce FSR 3.1, the latest iteration of their FidelityFX Super Resolution tech for game upscaling...

https://www.phoronix.com/news/AMD-FSR-3.1-Announced

[aufkrawall]

Hope a native and documented Vulkan demo & source code can help FOSS devs to match smooth present with HAGS on Windows.

Also the other changes are more than welcome. As much as DLSS is remarkable for being able to process image areas well with little available temporal history, I still hate its smearing and trailing artifacts. Instead of improving it, Nvidia also rather likes to ride the PR train with ray reconstruction, which, apart from RT details, often looks just terrible...

[theriddick]

The issue with FSR has always been excessive ghosting compared to DLSS and XeSS. The game that shows this issue the most is CP77, loads of videos on YT exampling what I mean.

[kiffmet]

It would be nice if VKD3D and DXVK could use this Vulkan version of FSR3 in order to inject it into games.

[SSJBurger]

The issue with FSR has always been excessive ghosting compared to DLSS and XeSS. The game that shows this issue the most is CP77, loads of videos on YT exampling what I mean.

That is mainly a CP77 issue as DLSS also has ghosting but at a very very minimal amount. Theres actually a video 4 days ago you can watch comparing them.

https://www.youtube.com/watch?v=CbJYtixMUgI

[aufkrawall]

The issue with FSR has always been excessive ghosting compared to DLSS and XeSS.

Can't be generalized like that. It has classic TAA ghosting issues, yes (also depends on the game), and they can also get more distinct with SSAO or RT effects that need heavy temporal smoothing.
But it doesn't "ghost" with objects moving in front of the skybox like birds or the lamp posts in CP77, where both XeSS and DLSS look terrible due to erroneous pixel clamping of the neural networks. XeSS also generally can't handle moving distant objects well and blurs and smears them even way more than DLSS.
Most of the reviews are unfortunately very bad in pointing out these issues and tend to ignore blurring issues of DLSS/XeSS over FSR's pixelation issues.

[RejectModernity]

It would be nice if VKD3D and DXVK could use this Vulkan version of FSR3 in order to inject it into games.

You can't inject it. Developers must wire it up by themselves.

[pWe00Iri3e7Z9lHOX2Qx]

Can someone remind me if there are any limitations for FSR 3.x for RDNA1 and RDNA2 generation cards?

[theriddick]

DLSS Ray Reconstruction is a pile of poop. At least until it gets remade because atm it blurs and smears everything.

[sobrus]

The issue with FSR has always been excessive ghosting compared to DLSS and XeSS. The game that shows this issue the most is CP77, loads of videos on YT exampling what I mean.

Both XeSS and DLSS are AI based solutions, while FSR probably uses classic algorithms. From my experience, no upscaling algorithm can match AI based ones, so worse FSR quality is to be expected. It may never be fixed, and I've heard that AMD is working on AI upscaler too.
But on the other hand, it's faster and supports more cards.

Thankfully Intel also implemented XeSS using DP4A instructions, and this is as good AI upscaler as RDNA2 can handle.
And it's not that much slower, while providing clearly superior IQ than FSR. Even if version using tensor cores on Arc is even better.
So now we have choice.