-
We’re currently investigating an issue related to the forum theme and styling that is impacting page layout and visual formatting. The problem has been identified, and we are actively working on a resolution. There is no impact to user data or functionality, this is strictly a front-end display issue. We’ll post an update once the fix has been deployed. Thanks for your patience while we get this sorted.
Discussion CNN DLSS and Transformer DLSS comparisons upscaling thread
- Thread starter poke01
- Start date
adroc_thurston
Diamond Member
It's incrementally better but also slower (can also be a mixed bag in places).
They threw parameter counts at the problem (but Catarzano was at least honest about that).
tl;dr ML upscaling plateaued.
They threw parameter counts at the problem (but Catarzano was at least honest about that).
tl;dr ML upscaling plateaued.
poke01
Diamond Member
I did various testing with the new model while not perfect is better overall than the CNN model. There are still improvements to be made, surfaces, vegetation etc need to be improved.
One thing I did notice though some games do get major improvements while others do not. Ghosting and motion clarity improvements being the main ones that the transformer model improves on.
Meteor Late
Senior member
Aka the main issue of TAA. Lol
poke01
Diamond Member
IMO this one of biggest upgrade, less VRAM usage because the performance mode DLSS4 looks just as good as quality DLSS3.8.
Should be great for lower end cards
Should be great for lower end cards
poke01
Diamond Member
coercitiv
Diamond Member
Does this reviewer not realize he's destroying the image quality for RT FOMO? It looks like playing games without prescription glasses.
poke01
Diamond Member
path tracing is a bit much for a 4060. yeah and look at the wiring/fence and edges are too soft, no good.
I would stick to performance mode as the minimum with PT off on a 4060 and get higher frames that way and retain some of the visual quality.
Meteor Late
Senior member
Well yeah, ultra performance is just too low of a resolution.
So then a tool like DLSSTweaks comes to the rescue, because the difference in base resolution reduction between Performance and Ultra Performance is just too big, with that tool one can just, for example, increase the base resolution for Ultra Performance mode. So if Performance has to scale 4x the resolution and Ultra performance is 9x, maybe with 6x scaling factor it would still look good and have a nice compromise in terms of performance and quality.
So then a tool like DLSSTweaks comes to the rescue, because the difference in base resolution reduction between Performance and Ultra Performance is just too big, with that tool one can just, for example, increase the base resolution for Ultra Performance mode. So if Performance has to scale 4x the resolution and Ultra performance is 9x, maybe with 6x scaling factor it would still look good and have a nice compromise in terms of performance and quality.
With 20 series now being over 6 years old, this is a big win for owners.
Meteor Late
Senior member
That generation is the perfect counterargument for the AMD fine wine thing, you are starting to see games that won't even launch with the 5700 XT due to Raytracing requirement, like Indiana Jones or Doom the Dark Ages, and some examples like Alan Wake 2 running much worse due to the use of mesh shaders. Meanwhile, RTX 2000 owners will be absolutely fine and with DLSS Transformer will get a bit more out of their card.
poke01
Diamond Member
You can actually run Indy on Linux on Vega 64/56 using software based Ray tracing. You can’t do that 5700 XT due to low compute.
So it’s goes like this RDNA3 > RDNA2 > Vega > RDNA1. RDNA1 is the worst AMD architecture ever.
coercitiv
Diamond Member
It's not similar. Over-sharpening was part of the criticism we had for both early DLSS and then early FSR, but nowadays the sharpening is considerably milder and often under user control. Sharpening will not solve AA artifacts.
This is the way, this is why I keep pushing for DLAA and FSR Native AA to be included in comparisons with upscaling examples, the improvement in image quality is something more people should be aware of.
Tup3x
Golden Member
It doesn't. Transformer model is much closer to what it should look like. Previous model boosted normal map detail way too much.Here transformer looks worse
SolidQ
Golden Member
Maybe, but skin detail in CNN more looks like in real life. I'm also showed to friends, what is better, they said all CNN.
Heartbreaker
Diamond Member
Is there a reason DLAA in DLSS 3.x is not compared, that already was the best option. Comparing to what DLAA in 3.x already beats seems pointless.
The real point is to see the supposed improvements over 3.x.
Heartbreaker
Diamond Member
Yes, I also heard the explanation, but I don't agree.
What it's doing may theoretically be better, but practically, the transformer model looks too smooth and plastic, and the old CNN model looks better.
Thunder 57
Diamond Member
RDNA1 beat Vega while using considerably less power. It also beat the 2060 and had more VRAM. It aged better until now.
Tup3x
Golden Member
In any case, previous ray reconstruction had that issue (IMO the left looks bad and also the faces looked bad in Cyberpunk). Now things that should be sharper generally look sharper without vaseline smearing. Nothing is perfect though.
Meteor Late
Senior member
The competition price point for the 5700 XT was mainly the 2060 super as HUB said, both 8GB cards.
There is no world where RDNA1 was going to age well with the HUGE technology deficit it had. 2070 super was 100$ higher priced, but the difference in technology feels much higher than that difference.