AMD FSR antialiasing discussion

[DegustatoR]

FSR doesnt reconstruct details and create new informationen.

It does actually. It paints over native image similarly to how SMAA / DLSS 1.0 did it. It is more complex than the former and less complex than the latter but it results in "reconstructed" image, with most of reconstruction happening on high contrast color transitions.

 

[Cyan]

dunno whether I understood it correctly or not, but..., for now it's not available on PS5, only on Xbox consoles?

if so there will be no competition performance wise this generation. If FSR manages to hold up fine image quality wise, even the Xbox Series S could give many surprises compared to the PS5, especially when it comes to performance and maybe image quality

 

[BRiT]

New dunno whether I understood it correctly or not, but..., for now it's not available on PS5, only on Xbox consoles?

No. Its for PCs right now.

 

[Cyan]

"In theory it could even be injected into a game by modders although AMD wasn't sure what would look correct this way, so your mileage may vary"

they are calling @iroboto out

 

[pcchen]

But i get that by just lowering the resolution. FSR doesnt reconstruct details and create new informationen. Buying a used 1080p display for gaming is more useful than using FSR to upscale from 1080p -> 2160p.

If it incorporate temporal information, then it does, to some extent. Before the super high pixel count image sensors and AI, people were using image sensors without low pass filter, taking multiple pictures with small offsets, to combine them into a single high resolution image (it's called "multi-frame super-resolution").
So, if the frame rate is high enough (60 fps is actually already quite high when motion is not large), a super resolution algorithm can reconstruct a higher resolution image from multiple frames. If the motion is large, then you probably won't notice the difference in resolution anyway.

 

[trinibwoy]

Watched the HWUB and DF videos. The consensus seems to be that 4K FSR ultra (1660p internal) works pretty well. This isn't surprising at all since 1660p without fancy upscaling looks pretty good too! At least from the couch on a 55" tv.

I think HWUB was a little premature in their stamp of approval based on just Godfall. The real test for FSR will be in games with fine detail (thin lines, wires etc) that benefit from temporal reconstruction. Those types of details just aren't present in Godfall. To be fair HWUB does mention this so hopefully they will revisit once more games are available.

The only real eye opening comparison was in the DF video of the character in a lacy dress. Holy crap. Huge difference between TAAU and FSR performance mode. That should convince anyone of the benefits of temporal data and it really does limit the potential reach of FSR 1.0 given most major engines have native support for DLSS or rolled their own temporal upscaler.

 

[Scott_Arm]

It has it's use. I can see myself using it in games where I want to maintain 140+ fps and make sure I'm cpu-limited to reduce input lag. Better than lowering render scale.

 

[Lurkmass]

Not sure why there is so much emphasis from DF trying to differentiate "upscaling" from "reconstruction" when their both indistinguishable from the point of view of the Nyquist theorem ...

 

[Dural]

those are the slides I was most interested in! Thx for sharing those slides Especially the 1440p numbers.

The future seems very interesting for this tech. However, the main downside imho would be if this technology only worked in new games.

If it could be forced into the GPU drivers to support our favourite games ever (Divinity Original Sin 2, Skyrim, etc etc) then I think there is no contest at all, it'd be the best upscaling technique to date hands down.

Hardware Unboxed review. @London Geezer they measure the image quality at 1440p too, which is what I also wanted to watch.

edit: according to them Terminator Resistance looks better with FSR on at ultra quality 1440p than native 1440p.

edit2: They also compare FSR not only between different native resolutions, including 1080p and 1440p, but mention GPUs such as the RX 570 4GB which they benchmarked with it as well. In addition they compare the FSR with typical upscale techniques -it wins every time-, also with Fidelity CAS where FSR again wins, and even with the best sharpening filters of adobe premiere , and surprisingly so they compare it to DLSS 1.0 -FSR wins again-, etc etc. And they even compare it with DLSS 2.0 and say it's very difficult, it's not a black or white thing, it's more complex than that, they say.

They explain why DLSS 2.0 is not necessarily better than FSR as of now, in some areas, while it is in others.

Of course they do, they hate Nvidia. I’d take anything they say with a grain of salt. My eyes tell me a much different tale than what they are trying to push. FSR is pretty bad in anything but ultra and even that sucks compared to other upscaling techniques.

 

[techuse]

Meh. What I expected. Better than DLSS 1. Worse than DLSS 2.

 

[MfA]

I think this is bad marketing, the effort and marketing would be better spent improving and promoting TAAU and maybe do some research for their own trained interpolater/clean up code which doesn't require MLP acceleration (I still suggest TSVQ).

This push actually undermines TAAU which is just a fucking terrible idea.

 

[Leoneazzurro5]

Watched the HWUB and DF videos. The consensus seems to be that 4K FSR ultra (1660p internal) works pretty well. This isn't surprising at all since 1660p without fancy upscaling looks pretty good too! At least from the couch on a 55" tv.

I think HWUB was a little premature in their stamp of approval based on just Godfall. The real test for FSR will be in games with fine detail (thin lines, wires etc) that benefit from temporal reconstruction. Those types of details just aren't present in Godfall. To be fair HWUB does mention this so hopefully they will revisit once more games are available.

The only real eye opening comparison was in the DF video of the character in a lacy dress. Holy crap. Huge difference between TAAU and FSR performance mode. That should convince anyone of the benefits of temporal data and it really does limit the potential reach of FSR 1.0 given most major engines have native support for DLSS or rolled their own temporal upscaler.

HWUB tested at least also Anno 1800, not just Godfall. The DF comparison between modes was made with performance setting (the one even AMD did not want their customer to use) and while there are some areas (transparency) looking bad, there are some points (the tree, the bricks on the wall) where the FSR shot looked better, at least to me. If engines have their own temporal upscaler, well, the question is why almost no one use it (except maybe the upcoming UE5). maybe because there are temporal artifacts? FSR is far from being a good solution (for now, future iterations maybe based on their patent and more may improve this), but in some cases it could be a good compromise, especially on dated cards.

 

[Leoneazzurro5]

This push actually undermines TAAU which is just a fucking terrible idea.

Not really, FSR in this iteration is only another tool that may or may not be used. It does not substitute TAAU or other temporal AA methods, and it is also said that it should work with existing AA (TAA) methods.

 

[CarstenS]

Most people who actually care about gaming do. The fact that it's 1060 which leads Steam h/w survey proves this without any doubt.

The 1060 with 6 GB was available for much of it's lifetime for 230-250 dollars/Euro (here: with VAT, so basically equal), sometimes even less. And this does not even take into account the 3 GByte version.
So, thanks for making my point.

Attached for your convenience a screenshot from our local price search engine of a regular Zotac model. https://geizhals.de/zotac-geforce-gtx-1060-zt-p10620a-10m-a1942856.html

 

[CarstenS]

No. Its for PCs right now.

It is, but apparently, Microsoft was closely involved :

[Timestamped link to GDC-2019-Video about Super Resolution]
and some of the developer quotes on FSR seem to indicate, it's being implemented on consoles as we speak.
edit:
Here's the quote with my bolding:
"Implementing AMD Fidelity FX Super Resolution into The Riftbreaker was one of the smoothest technology integrations that we have ever done. Thanks to AMD's "no black box" philosophy we were able to very easily implement it on both PC and consoles and tweak it for the specifics of our game engine.

FSR is a cutting-edge upscaling solution that's available on all modern graphics hardware and combines superior quality with blazing fast performance.

Pawel Lekki, COO, Exor Studios"
https://gpuopen.com/fidelityfx-superresolution/#performance

 

[snc]

If you use 1440P base resolution + FSR, you get a higher resolution. If you have native 1440p, and output 1440p, then you can't compare the images without downsampling the 4K FSR back to 1440p.

If you decide to measure 1440p ->4K FSR, vs 1440p outputting at 4K, then you've chosen bilinear upsampling for the native image, so in theory you could also choose TAA Upsampling for instance.

I'm not sure a proper comparison can be done.

You'll lose the frame rate advantage of comparing 1440p vs 1440pFSR.

Technically, according to these graphs, to use Ultra Quality, both consoles will likely incur about a 1ms penalty for processing.

You will lose even more perf comparing 1440p i internal to 4k and 1440p dlss2 but you will get better quality, question is how much better looks fsr from 1440p to 4k vs 1440p (or similar resolution)

 

[MfA]

Not really, FSR in this iteration is only another tool that may or may not be used. It does not substitute TAAU or other temporal AA methods, and it is also said that it should work with existing AA (TAA) methods.

FSR is a purely management driven response to DLSS, a bad response. TAAU has some problems, but it's still the most immediate answer AMD has. NVIDIA will wink wink nudge nudge TAAU to death for PC if they can get away with it, if AMD refuses to actively push it it will just continue to be an anonymous feature only turned on for consoles.

 

[Leoneazzurro5]

FSR is a purely management driven response to DLSS, a bad response. TAAU has some problems, but it's still the most immediate answer AMD has. NVIDIA will wink wink nudge nudge TAAU to death for PC if they can get away with it, if AMD refuses to actively push it it will just continue to be an anonymous feature only turned on for consoles.

TAAU is something in control of developers, not AMD. I agree that something like UE5's TSR is probably the future, but this is not something AMD can force on software houses. And IIRC AMD helped in optimizing UE5 for AMD GPUs, including TSR..

 

[troyan]

If it incorporate temporal information, then it does, to some extent. Before the super high pixel count image sensors and AI, people were using image sensors without low pass filter, taking multiple pictures with small offsets, to combine them into a single high resolution image (it's called "multi-frame super-resolution").
So, if the frame rate is high enough (60 fps is actually already quite high when motion is not large), a super resolution algorithm can reconstruct a higher resolution image from multiple frames. If the motion is large, then you probably won't notice the difference in resolution anyway.

And Epic's current TAAU is actually really good at it. But FSR doesnt do anything for reconstructing images, it is just upscaling a low res image.

 

[Rootax]

And Epic's current TAAU is actually really good at it. But FSR doesnt do anything for reconstructing images, it is just upscaling a low res image.

I really don't get the différence between upscaling and reconstructing for the end user. From a tech POV, yes I get it, but if the PQ is here in the end, it doesn't really mater ?