• Welcome to TechPowerUp Forums, Guest! Please check out our forum guidelines for info related to our community.

AMD FSR FidelityFX Super Resolution Quality & Performance

Vya Domus said:
It's better to just look at AMD's own patent for this :
View attachment 205226

It's constructing higher resolution images that fit the extracted features from lower resolution input images, hence it adds additional detail or new information or whatever you want to call it. If you want to maintain this stance that it is not adding detail then you have apply this logic to everything else, including DLSS.

Even a straight forward linear interpolation adds new information that wasn't present in the initial image, that's the point of these algorithms otherwise there would be no use for such a thing, it might be rubbish but that's how it works.
Click to expand...
Adding more information does not equal detail.
But I'm not gonna keep beating this anymore.
 
Camm said:
Reshade *should* be able to inject FSR by the looks. Wouldn't be surprised if Reshade gets support once its open sourced / when AMD open sources it we may get a 'freestyle' style shader option for AMD (this is entirely scuttlebutt on my behalf).
Click to expand...
It basically already does AMD's contrast adaptive sharpen which is a big portion of FSR was already ported to Reshade for awhile now. I literally already showed a link to a example of it in this thread earlier already. Even at 720p it's readily noticeable. In fact I did a bit of testing and compared CAS enabled with texture quality reduced by half and it subjectively was overall better in quality. I'm not entirely how big the ramifications of that are at higher resolutions I tested it at 720p, but I imagine it's pretty big. I might have to revisit that scene and try it again at a higher resolution.

The big difference is the shading and lighting greatly improved. I've been trying to improve my configuration further experimenting with it I want to try to get improved results at even lower resolutions like 576p because it'll translate into bigger gains at higher resolutions. I've found that as I improve my post process techniques I've been able to generally reduce the CPU/GPU overhead as well required to get the type of results I want and with less unnatural jarring side effects at the same time.
 
 
Auer said:
Adding more information does not equal detail.
Click to expand...
This makes zero sense.

You’re telling me this piece of software adds what then if not detail aka information ? Noise ?
 
It sure looks like it adds information...

Default...
De)fault.jpg


AMD based contrast adaptive sharpen at half the texture resolution....
Re)shade.jpg



You tell me which looks better default or contrast adaptive sharpen...to me subjectively the default looks overall worse sure there are differences you can point to, but overall which would you pick. I can point out more upside than downside with the CAS options...there is also this...

CAS with the shame texture resolution as the default...if you really think it doesn't look any better or add any information I think you're a little bit crazy. I mean it doesn't upscale the pixel size at least not on it's own with just CAS, but it absolutely improves the overall colors space to provide richer scene detail. I also don't see any obvious halo ringing effect. I mean you can pretend because it's not DLSS it does nothing and sucks, but in reality it does a lot for free for a lot of people and is perceptibly a winning improvement and no janky blur or in motion temporal nausea at the same time.
Re)shade HQ.jpg
 
When they say DLSS is adding more details, what most people say is details that are too small and not on the right grid to appear in the native image or the lower res image before upscaling.

DLSS using Temporal AA and a method of switching the grid each frame, is able to display these details that otherwise wouldn't appear.

FSR can only enhance what it's already there but lets say there is a cable far away that is less than 1 pixel wide at 1440p, the lower res version will have no data to extract from it.

But FSR add information to a frame by using their extrapolation and sharpening technique. So both of you are right on your own.
 
Give it time and I'm sure AMD and developers will improve upon the different aspects that make up FSR as well as add other elements into the fold. It's very open in nature you can already combine TAA with FSR in practice you could do it right now with Rift Breaker injecting TAA with Reshade alongside the FSR nothing is stopping you.

FSR improves performance and looks good, but DLSS uses temporal AA yeah well FSR could too. It could use machine learning as well technically, but doesn't have dedicated hardware for it right now. Still AMD could add ML to their CPU's or APU's or whatever and do just that. It's a pretty good start in either case and seems like a better start than DLSS 1.0 was and is already being pretty widely supported for something that "sucks" at information of things according to some green goblin fanatics. Does FSR hurt your feels that much do you want your money back!!? By the end of the year I wouldn't be shocked if FSR is more widely supported than DLSS.
 
InVasMani said:
It sure looks like it adds information...


You tell me which looks better default or contrast adaptive sharpen
Click to expand...

I'm not commenting per say (on a technical level any upscale if it isn't integer is adding or subtracting info so take that as you will).

But an interesting test here would be on a 1440p panel, take a screenshot at 1440p native, 4k DSR, 4k DSR FSR, and 2954x1662 DSR then compare data.

You have

A: A native res control image
B: A fully rendered upscaled image which should be integer perfect
(These are your control images)
C: A integer upscaled FSR image
D: A fully rendered FSR native res image.

Between the 4 there, you should be able to tell if FSR adds information. If it doesn't, 4k DSR FSR, and 2954x1662 DSR should look identical minus sharpening.
 
It may be that Nvidia's DLSS is better, but as far as I can tell, FSR is definitely good enough. Of course, turning off ray tracing, using a lower resolution, and using lower quality settings are still going to be the options chosen by competitive gamers; their interest will be in things like Nvidia's Reflex technology. But not every video game is Fortnite, people play some types of single-player games looking for an immersive experience, and the highest possible video quality serves that.
 
The focus should really be on how can we and developers best take advantage of FSR to improve the free give AMD gave us. Clearly there are ways to improve it. Technically you don't need image to improve upscale and add detail there are other manners of adding detail at the pixel level. The pixel shaders themselves adapt and modify the images themselves so they should be able to do so form the previous shader in a chain of pixel shader configurations in theory if I'm not mistaken. I would think technically the base image would be enough provided your clever enough to manipulate them well in series.
 
quadibloc said:
It may be that Nvidia's DLSS is better, but as far as I can tell, FSR is definitely good enough. Of course, turning off ray tracing, using a lower resolution, and using lower quality settings are still going to be the options chosen by competitive gamers; their interest will be in things like Nvidia's Reflex technology. But not every video game is Fortnite, people play some types of single-player games looking for an immersive experience, and the highest possible video quality serves that.
Click to expand...
It's all depend on what you judge the upscaling method.

If you judge an upscaling method only by comparing 400x zoomed subpixel detail on screenshot. Yes indeed, if that is the only factor.

But there are way more, Performance, ease of implementation, Game supporting it, Hardware supporting it, Behavior with effect and post-effect, how it react in movement at low or high fps. is it good with low res or require higher res to be effective.

It still a bit soon to determine what would be the winning upscaling method.
 
Last edited:
I don't think FSR is perfect or perfected yet, but I think AMD defiantly has knocked it out of the park with FSR just the same with the impact it'll have on many systems around the globe.
 
InVasMani said:
Why is Myst on the list was there a remake or something!!!?
Click to expand...
Welp, I think the fact that FSR is extremely easy to add to a game, motivated devs to include it into a rather old game.
 
DF's Alex shilling way too hard, destroying his own credibility in the process. Amusing to watch.:laugh:
 
Punkenjoy said:
FSR can only enhance what it's already there but lets say there is a cable far away that is less than 1 pixel wide at 1440p, the lower res version will have no data to extract from it.
Click to expand...
This is true indeed. But if it is less than 1 pixel at 1440p, it is just 1 pixel at 4k, something barely visible anyway.

On top of it, "ultra" quality renders at resolution higher than 1440p (which is 2.2 times less pixels than 4k), it renders at resolution that is only 1.3 times behind.

quadibloc said:
their interest will be in things like Nvidia's Reflex
Click to expand...
This is outright insulting to read on a tech forum.
NV's reflex is copypasta of AMD's Anti-Lag, on top of having nothing to do with upscaling.
 
Console Tech

www.techpowerup.com

AMD FSR FidelityFX Super Resolution is Coming to Xbox Consoles

Just a few days ago, we have reviewed AMD's FidelityFX Super Resolution technology, which represents an answer to NVIDIA's Deep Learning Super Sampling technology used to upscale images t certain resolutions. As the review predicted, AMD's presence in consoles like PlayStation 5 and Xbox Series...
www.techpowerup.com www.techpowerup.com
 
Contains a lengthy discussion on the controversy of the DF video, with some interesting points made.

 
They still omit the big downside of TAA. TAA can interfere with many effects due to his temporal nature and not every one want to have to deal with that. FSR is intended to be a easy to use upscaler that devs don't have to worry to much about it and make it in the game easily.

I don't think big AAA studios, except being paid by AMD would use that if they can have TAA Upscaler already covered. The whole point of AMD was to have a easy solution that dev can implement easily and get some mind share. The way FSR work, it can be put into any 3d game easily without artefact.

So that is back to the point on what you judge your Upscaler. If it's only image quality, FSR would probably lose, but if it's just to provide the ability to do resolution scaling in a game that don't have it. Just getting that would be great.

But would i prefer UE 5 TAA U ? probably, but far from all game would be make with that. There would probably be UE 4 engine game for year and their TAA U isn't working perfectly.
 
Punkenjoy said:
I don't think big AAA studios, except being paid by AMD would use that if they can have TAA Upscaler already covered. The whole point of AMD was to have a easy solution that dev can implement easily and get some mind share. The way FSR work, it can be put into any 3d game easily without artefact.
Click to expand...
That's pretty much what AMD RIS did.
 
InVasMani said:
Give it time and I'm sure AMD and developers will improve upon the different aspects that make up FSR as well as add other elements into the fold. It's very open in nature you can already combine TAA with FSR in practice you could do it right now with Rift Breaker injecting TAA with Reshade alongside the FSR nothing is stopping you.

FSR improves performance and looks good, but DLSS uses temporal AA yeah well FSR could too. It could use machine learning as well technically, but doesn't have dedicated hardware for it right now. Still AMD could add ML to their CPU's or APU's or whatever and do just that. It's a pretty good start in either case and seems like a better start than DLSS 1.0 was and is already being pretty widely supported for something that "sucks" at information of things according to some green goblin fanatics. Does FSR hurt your feels that much do you want your money back!!? By the end of the year I wouldn't be shocked if FSR is more widely supported than DLSS.
Click to expand...
In my opinion, FSR or DLSS each have their pros and cons. Image quality wise, DLSS seems to have the edge because of the complexity of the implementation. Ultimately, both methods give us improved performance without sacrificing too much in terms of visuals, which I think is a win. Even for people who own a shiny new RTX 3000 series or the older 2000 series, FSR is a good fallback in case a game doesn't support DLSS. As an owner of a RTX 3000 series card, I am not complaining. And knowing Nvidia, at some point they will drop support for older cards, which by then if FSR is still around, it could benefit even current Ampere users in the long run. After all, using FSR is better than lowering the native resolution or graphical settings, if your GPU is struggling.
 
How do you know what's sharpened up and what isn't ?
Click to expand...
Auer said:
This is basically just AMD's existing CAS, part of the whole FSR deal.
It's always there.
Click to expand...
Indeed, it's how it works, if it looks sharper than native, I'm going to go with it was sharpened.
 
Resident Evil Village has FSR already.

R5 2600, 16GB, RX 560 4GB - 1920x1080 - Preset Performance Priority (Low) - Dimitrescu Main Hall
FSR Off - 53fps
FSR Ultra - 69fps
FSR Quality - 79fps
FSR Balanced - 87fps => I can increase graphic to medium & got 65fps.
FSR Performance - 96fps => Feeling like playing SNES emulator with 2D scaling.

Previously, I need to use resolution 1366x768px on medium graphic.
But now, I can play at Full HD medium easily, upgraded from 45-55fps to 65-85fps using FSR Balanced.
The quality of FSR image looking good. Better than 1366x768 and better than low quality graphic at full hd.

1626777370385.png
 
Last edited:
You must log in or register to reply here.
Back
Top