AMD FSR antialiasing discussion

  • Thread starter Deleted member 90741
  • Start date
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 ?

FSR doesnt deliver the PQ from lower base resolutions. Maybe it is acceptable when you cant compare it to other techniques. For example DLSS in Watch Dogs Legions:
and FSR in Godfall:

Performance setting uses only 1/4 of the pixel of the native image. DLSS can reconstruct a lot of information, FSR is just upscaling a low res low information 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 ?

Theres no difference to the end user so it’s really an academic distinction. I think the point is that a spatial upscaler doesn’t even attempt to fill in detail that is missing from the frame. It upscales what’s there and tries to not make it look blurry in the process. Whereas temporal techniques have way more data that can be used to actually fill in detail missing from the current frame. Not to mention you can use the same data to apply antialiasing at the same time.

It’s really hard to see why a developer would choose TAA + FSR over TAAU.
 
FSR doesnt deliver the PQ from lower base resolutions. Maybe it is acceptable when you cant compare it to other techniques. For example DLSS in Watch Dogs Legions:
and FSR in Godfall:

Performance setting uses only 1/4 of the pixel of the native image. DLSS can reconstruct a lot of information, FSR is just upscaling a low res low information image.


Why focus on Performance setting when quality and ultra quality is where the sweet PQ/performances spot is ?
 
Quality and Ultra Quality are not useful for people with older GPUs. FSR doesnt solve existence problems for them. Ultra Quality in 4K is 1660p as the base resolution. And this is the only quality setting from FSR which results in accectable image quality. Dont know how many users with older GPUs (5700XT or RTX2070 for example) get enough frames in 1660p...
 
How good is a temporal component with ever older temporal samples? Are they really temporal or are they rather successive? What I mean is, with low fps rates in the 20s, your first sample from, say, 8 frames ago is notably older (temporally speaking) than if you were to apply this technique to a setup running at 120 fps. Is there a time-based cutoff?
 
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.

Well that’s the thing. AMD acknowledges that FSR is no good upscaling 1080p to 4K while we see TAAU do an acceptable job doing just that. TAAU also runs on older hardware. So in the end why not just go with TAAU which has a much more useful range of scaling ratios?

The FSR advantage of being easier to implement in a game doesn’t seem to be that helpful given TAA is basically in everything anyway. So FSR is really only easier to add to a game that doesn’t use TAA and that’s a shrinking population.
 
TAAU is something in control of developers, not AMD.
So is their decision to include FSR. AMD is using money and marketing to nudge them to adopt FSR ... no one is using money and marketing to push TAAU on PCs, not even Epic. It will get nudge nudge wink winked to death on PC while being the more competitive option.
 
Quality and Ultra Quality are not useful for people with older GPUs. FSR doesnt solve existence problems for them. Ultra Quality in 4K is 1660p as the base resolution. And this is the only quality setting from FSR which results in accectable image quality. Dont know how many users with older GPUs (5700XT or RTX2070 for example) get enough frames in 1660p...
The 5700 XT averages 52.6 FPS at 4K in TechpowerUP's suite. So it could be useful for hitting 60 FPS more consistently.

https://www.techpowerup.com/review/nvidia-geforce-rtx-3070-ti-founders-edition/27.html
 
How good is a temporal component with ever older temporal samples? Are they really temporal or are they rather successive? What I mean is, with low fps rates in the 20s, your first sample from, say, 8 frames ago is notably older (temporally speaking) than if you were to apply this technique to a setup running at 120 fps. Is there a time-based cutoff?

TAA gives more weight to the newest frame so the contribution from the oldest frame amortizes down to zero over time. This works best when you have a high frame rate.

In a low frame rate scenario where camera or object movement can significantly change the color of a pixel over a few frames TAA should detect this change and essentially throw away all the old data for that pixel and start accumulating from scratch.

This paper gives a good overview. http://behindthepixels.io/assets/files/TemporalAA.pdf
 
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.
There are clear and obvious IQ degradations in Godfall even on UQ preset, this is shown in pretty much every other publication on FSR so far. HUB are just being themselves again.
The games which fare best are Kingshunt and Terminator - mostly because they are very low detailed and blurred by post processing to boot.
Riftbreaker is the best case IMO - it is both highly detailed and doesn't loose much from FSR's UQ. But still you can easily spot the difference between native and UQ even in YouTube videos (compression and all).
So it really depends on how you present this. It's not close to native or to DLSS or even TAAU. It is good in comparison to upscaling with your monitor.

The 1060 with 6 GB was available for much of it's lifetime for 230-250 dollars/Euro (here: with VAT, so basically equal)
The card launched at $300 and was $250 only for bare minimum models.
In any case this is much closer to $300 than to 1050 or 1050Ti price.

Does it? Put the same amount of sharpening on TAAU output and compare scenes with temporal aliasing.
 
Last edited:
Well that’s the thing. AMD acknowledges that FSR is no good upscaling 1080p to 4K while we see TAAU do an acceptable job doing just that. TAAU also runs on older hardware. So in the end why not just go with TAAU which has a much more useful range of scaling ratios?

The FSR advantage of being easier to implement in a game doesn’t seem to be that helpful given TAA is basically in everything anyway. So FSR is really only easier to add to a game that doesn’t use TAA and that’s a shrinking population.

Some screenshots above regarding Terminator Resistance prove that TAAU at least in some instances does a worse job.
In any case, personally I think good implementation of image reconstruction techniques like TSR (which is also hardware agnositc) are the future, question is how many can/want implement them in an effective way.
 
Last edited:
So is their decision to include FSR. AMD is using money and marketing to nudge them to adopt FSR ... no one is using money and marketing to push TAAU on PCs, not even Epic. It will get nudge nudge wink winked to death on PC while being the more competitive option.

If AMD helped Epic to optimize TSR and UE5, they ARE investing some time (and thus money) in such techniques. But yes, this is mainly a marketing move.
 
Is there a time-based cutoff?
No, frames are being merged recursively.
TAA usually works with just 2 consequent frames - current low res and already reconstructed high res previous one (the history buffer), low res one is being upscaled to higher resolution first (that's why spatial upscaling is important for TAAU and DLSS too), then it's being merged with warped history.
For TAAU, exponential moving average is used to blend between frames hence far away frames don't contribute too much (but if there is some error, it will stay here for a bunch of frames that's why you can see ghosting with TAA).
 
Does it? Put the same amount of sharpening on TAAU output and compare scenes with temporal aliasing.
PCGH here is unfortunately wrong about Terminator Resistance's res slider using TAA Upsample. Here if you inject Unreal Engine Unlocker into the game and set the res slider below 100, you can see that TAA Upsampling is set to 0:
temrinatorlrk58.jpg

So their image comparison is FSR against the standard non-taa upscale.
 
Techpowerup comparison tool was the eye opener.

Go fullscreen, pick the quality, use the mouse scroll wheel to zoom in and out and side-to-side, and judge for yourself.

Having done that, It was a stark contrast when first seeing the DF video and then seeing the high-res results for myself on techpowerup. The end result doesn't care if there are temporal alternatives out there. I was impressed. It is legitimately a good tool with its own advantages and should exist on its own merit.
 
PCGH here is unfortunately wrong about Terminator Resistance's res slider using TAA Upsample. Here if you inject Unreal Engine Unlocker into the game and set the res slider below 100, you can see that TAA Upsampling is set to 0:

So their image comparison is FSR against the standard non-taa upscale.
Yeah, I thought it looked way too blurry to be TAAU.

This is how much UE4's temporal upsampling improves image quality when scaling from 720p to 1440p. https://imgsli.com/NTg1NDg
 
Well that’s the thing. AMD acknowledges that FSR is no good upscaling 1080p to 4K while we see TAAU do an acceptable job doing just that. TAAU also runs on older hardware. So in the end why not just go with TAAU which has a much more useful range of scaling ratios?

Because there are developers who don't ever want to implement any temporal sampling ...

The FSR advantage of being easier to implement in a game doesn’t seem to be that helpful given TAA is basically in everything anyway. So FSR is really only easier to add to a game that doesn’t use TAA and that’s a shrinking population.

TAA is only common in console centric AAA games. Many small developers don't ever implement TAA. Anno 1800 is an FSR supported game that doesn't even have TAA either. Godot will never have any temporal sampling techniques because of the lead developer's distaste for it's implementation cost. Why deny the advantage of implementation simplicity when some developers feel that it's very much real to them ?

What DF did by comparing FSR to one of the state of the art temporal upsamplers was somewhat disingenuous since Epic Games have dozens of engineers to optimize the quality of their technique when a lot of studios don't even come close with their own existing TAA solutions ...
 
Back
Top