AMD FidelityFX Super Resolution 2.0 Source Code Published

[jrch2k8]

This is not what professional reviewers have concluded:

It's decent, but it's definitely not better.

is better for 99% of the people on most common situations tho, nobody is going to play and zoom 3 freaking kms away just to see if a pixel is missing in the next block building, realistically speaking

[luno]

From experience I don’t think it looks better. It competes with Nvidia’s Image Scaling along with every other traditional upscaler that’s in game engines. For example UE5 which has Temporal Super Resolution. I’m certain Intel’s XeSS, Nvidia’s DLSS will continue to be in leagues of their own. I know Microsoft/Sony are also working on their own AI upscalers. FaceBook already has their own called neural supersampling. AMD has been working on an AI solution since the release of the Radeon VII.

even Apple has Upscaler https://developer.apple.com/documentation/metalfx

[birdie]

is better for 99% of the people on most common situations tho, nobody is going to play and zoom 3 freaking kms away just to see if a pixel is missing in the next block building, realistically speaking

You haven't watched the entire video, have you? There are instances where FSR 2.0 is a pain to look at. It's not a long video, please devote your time to it, and stop making conclusions based on the last 30 seconds.

[WannaBeOCer]

even Apple has Upscaler https://developer.apple.com/documentation/metalfx

Didn’t mention it since it’s not an AI. It’s jut another traditional upscaler which you can find in practically every game engine.

[Mahboi]

Impressive.

A shame they do not accept PRs, it seems to be on an automatic closure policy right now.

I hope that their implementation will become a standard (until FSR 3.0 at least) and that tools like Godot Engine can take it for themselves. They've already done it for FSR 1.0, and someone even made a better version of it. Hopefully this goes on.

Edit: they DO accept PRs and their bot is going to chill, apparently.

[jrch2k8]

You haven't watched the entire video, have you? There are instances where FSR 2.0 is a pain to look at. It's not a long video, please devote your time to it, and stop making conclusions based on the last 30 seconds.

watched fully at 4k and i have played the damn thing, if the video dude don't zoom, freeze, capture specific sequences and loop them and give an explanation is damn fucking hard to notice, i actually never did during gameplay, my point is for a regular user is barely noticeable at best and it have to be looking for it.

Not saying technically speaking is perfect, im saying is good enough for 99% of us mere mortals

[CochainComplex]

watched fully at 4k and i have played the damn thing, if the video dude don't zoom, freeze, capture specific sequences and loop them and give an explanation is damn fucking hard to notice, i actually never did during gameplay, my point is for a regular user is barely noticeable at best and it have to be looking for it.

Not saying technically speaking is perfect, im saying is good enough for 99% of us mere mortals

I also watch it on 4k. Havent played it. Yes if you really have a close look DLSS is better. But in 99% during gaming you might not notice a huge difference. The "smearing" of some water surfaces might be a more obviouse downsides. But really AMD did a good job.

[FedNat]

Didn’t mention it since it’s not an AI. It’s jut another traditional upscaler which you can find in practically every game engine.

AI is not a magic bullet and FSR 2.0 clearly shows that's not needed to have good results on upscaling

[WannaBeOCer]

AI is not a magic bullet and FSR 2.0 clearly shows that's not needed to have good results on upscaling

FSR 2.0 doesn’t prove anything. It has the same draw backs of all the other temporal upscalers I’ve seen. Only thing AMD did right was make it open source. While all the other upscalers are game engine specific. DLSS and other AI upscalers that utilize inference dedicated hardware like tensor cores/avx-512 or matrix cores will also run faster which is why we see a 5-8% increase using DLSS compared to FSR on the same GPU. Isn’t that the main purpose of using an upscaler? To gain more performance instead of buying more expensive hardware?

[brucethemoose]

I wonder if this can be modified to not need a depth buffer, so it might be applicable to video files with automatically generated motion vectors...

Depth buffers can be generated, see: https://github.com/baowenbo/DAIN

This could definitely be ported to VapourSynth. But doing it in real time... well that's another matter.

Maybe it would still be OK in mpv if you gave it a dummy or really basic depth buffer? Or maybe there is a toggle, I havent looked at the code yet.