AMD FSR antialiasing discussion

  • Thread starter Deleted member 90741
  • Start date
Also strange how they downplay DLSS at lower resolutions like 1080p, where the differences between FSR and DLSS are much more obvious, with the narrative that both upscaling techniques hurt image quality too much to be useful. When infact their own data demonstrates that DLSS at 1080p looks better than native 1080p lol: https://imgsli.com/NjE4NDI

How is that even native? Both are obviously TAA .
 
Also strange how they downplay DLSS at lower resolutions like 1080p, where the differences between FSR and DLSS are much more obvious, with the narrative that both upscaling techniques hurt image quality too much to be useful. When infact their own data demonstrates that DLSS at 1080p looks better than native 1080p lol: https://imgsli.com/NjE4NDI

DLSS will often look great in stills especially when reconstructing fine detail. To be fair HWUB's opinion was based on the examples in motion. I thought they did a very good job demonstrating motion artifacts ghosting, blurring etc.

Still screenshots are fine for FSR but aren't adequate for a proper discussion of TAA and DLSS.
 
In Avengers the TAA is dimming the brightness in movement which results in better AA quality. But this is far away from the native non TAA picture and blurs minimal the image in 4K. Using a lower resolution will only amplify the problem. DLSS is not as stable but overall produces a slighty better picture than TAA.

That is one of the problems of FSR. It inherit every problem of TAA and will make it worse.
 
How is that even native? Both are obviously TAA .

I think "native" is an ok descriptor. People tend to just mean "not upscaled." in the vid they complain about how bad the TAA implementation is in the avengers. Not sure if TAA can be turned off in that game. Would be nice to see TAA off as part of the comparisons.
 
See through speaker grill vs non see through :LOL:

Really though, this is a good example to use to give feedback to Nvidia so they can work on this stuff.

Is the full image for this comparison there? I'd like to see it within the context of the entire frame.

its on the video posred by @troyan

seems the grille is supposedly to be semi transparent. where the holes allows you to see the insides. DLSS simply make the holes black. While in native render, it didnt properly make the insides of the speaker darker
 
its on the video posred by @troyan

seems the grille is supposedly to be semi transparent. where the holes allows you to see the insides. DLSS simply make the holes black. While in native render, it didnt properly make the insides of the speaker darker
Oh ok yea, I just checked the video and it definitely isn't just some small background detail that was greatly zoomed it. It's a pretty considerable chunk of the frame to be completely missing that detail.

Weird considering how well DLSS handles fine details usually.
 
Yeah, agreed. That's a bit weird. Maybe looking for orderly structures a bit too heavily weighted?

Now, I'm not the biggest fan of DLSS myself, but in this case, I have two remarks. Regarding the blotchy pattern on the side of the speaker, this changes with head position (which is cut from the frame) and it being quite a bit behind the depth of the head, this should not be the case IMHO; so a game artifact?

Second: This is a 400 % zoomed in shot. It's not nearly that big in the original video. It's not a tiny minute detail of the background either. Judging from the video part that starts at 1:20, I could imagine here's some Mip-LOD at work, that's missing from/missed by DLSS.

And third. The bigger picture (a second or so after the shots from above) at 400% zoom already.
upload_2021-7-20_7-33-57.png
 
I think "native" is an ok descriptor. People tend to just mean "not upscaled." in the vid they complain about how bad the TAA implementation is in the avengers. Not sure if TAA can be turned off in that game. Would be nice to see TAA off as part of the comparisons.
For resolution, it is.

For AA and upscaling images there always should be additional information on what AA & upscale is used as well.

Native with TAA off and with >64xSSAA would be nice. (give base frame which is used for TAA and a target quality.)
 
Yeah, agreed. That's a bit weird. Maybe looking for orderly structures a bit too heavily weighted?

Now, I'm not the biggest fan of DLSS myself, but in this case, I have two remarks. Regarding the blotchy pattern on the side of the speaker, this changes with head position (which is cut from the frame) and it being quite a bit behind the depth of the head, this should not be the case IMHO; so a game artifact?

Second: This is a 400 % zoomed in shot. It's not nearly that big in the original video. It's not a tiny minute detail of the background either. Judging from the video part that starts at 1:20, I could imagine here's some Mip-LOD at work, that's missing from/missed by DLSS.

And third. The bigger picture (a second or so after the shots from above) at 400% zoom already.
View attachment 5747

Maybe DLSS trained to makes things looks proper?

The mesh of DLSS is actually looks more realistic/proper than native. As in reality the insides (cones) would be pretty dark/hard to see.

Same with the hair. The DLSS have much better hairs, including the tiny strands
 
the blotchy pattern on the side of the speaker also changed a bit in DLSS
The blotchy pattern is likely dithering to hide transition between different geometry LOD levels.
Looks like that's just another game where devs have not adjusted geometry LODs to account for lower internal resolution with DLSS or FSR, could have been easily checked by switching to FSR Quality with the same internal resolution as DLSS Quality.
 
Maybe DLSS trained to makes things looks proper?
It'd be nice if examples like this were presented during developer interviews to see how close the visual representations are from their original intent (assuming native may or may not be ideal).
 
seems DLSS uses way too much imagination on the speakers

View attachment 5746

EDIT:

the blotchy pattern on the side of the speaker also changed a bit in DLSS

Yeah, agreed. That's a bit weird. Maybe looking for orderly structures a bit too heavily weighted?

Now, I'm not the biggest fan of DLSS myself, but in this case, I have two remarks. Regarding the blotchy pattern on the side of the speaker, this changes with head position (which is cut from the frame) and it being quite a bit behind the depth of the head, this should not be the case IMHO; so a game artifact?

Second: This is a 400 % zoomed in shot. It's not nearly that big in the original video. It's not a tiny minute detail of the background either. Judging from the video part that starts at 1:20, I could imagine here's some Mip-LOD at work, that's missing from/missed by DLSS.

And third. The bigger picture (a second or so after the shots from above) at 400% zoom already.
View attachment 5747

That's a pretty funny omission and I agree it actually looks better in this case but could obviously easily look worse under slightly different circumstances. That aside though, there are some very clear wins for DLSS in that video even over native with SMAA. That closeup of Thors torso and beard in literally night and day.
 

That's great news and a great feather in FSR's cap IMO. The one area that REV falls short in vs the console versions (aside from the DRM issue which is being resolved) is in it's checkerboard upscaling technique which I understand is pretty poor vs the console version. I would hope that FSR can do a much better job and at least equalise things in that regard, if not give the PC version the advantage. It'd be nice to see Alex do an update to the current REV face off once FSR is in place and the DRM issue has been patched out.
 
Back
Top