Frenetic Pony
Veteran
That's really what I'm waiting for. Does it work with TAA? DLSS 1.0s problem wasn't that it didn't use TAA, it was that it didn't work with TAA. If this works with TAA while being a spatial filter as well, that's potentially pretty cool, and much easier to integrate into any given game engine (a goal apparently). In fact in some ways it's more useful. I can see a lot of triple a game engines using the spatial filter to upsample new screen areas before TAA kicks in next frame, removing the undersampled and jaggy look there. It's also potentially useful for upsamping smooth reflective surfaces.
Buuuut, where's the independent video comparison? Only relevant if it looks decent at all.
Rootax
Veteran
Anandtech seems to confirm that. Not really a surprise.
I'd still wait for details before jumping to the conclusions.
That patent from which the thread started was about spatial ML reconstruction after all.
That patent from which the thread started was about spatial ML reconstruction after all.
D
Deleted member 13524
Guest
Yup, the patent described linear and non-linear upscaling, but no motion vectors.
Perhaps it can be used with TAA.
It might also be less dependent on framerate.
Though I do admit that quality setting screenshot doesn't look good.
hope this is just for the start. Future update might bring a better technology ( or not )
Because FSR has much higher input resolution to upscale from at 4K than 1440p.
Also, do you have a 4K monitor? Otherwise you are not going to see huge differences in image quality. Especially not on a YouTube video.
I wonder if running FSR on monitor native res to some DSR/VSR res and downscaling back to native would look better than just native res rendering.
With DLSS this is true at least so this maybe one interesting application of FSR even if its upscaling quality won't be that good.
With DLSS this is true at least so this maybe one interesting application of FSR even if its upscaling quality won't be that good.
D
Deleted member 13524
Guest
Epic's Temporal Super Resolution was made in collaboration with AMD according to the UE5 documentation, so it's probably FSR + Epic's temporal reconstruction.
It does look like AMD purposely left temporal data out of FSR to guarantee a higher level of compatibility across engines, and boost adoption that way.
There are 3 tiers of IQ/Performance: Performance, Quality and Ultra Quality.
The GTX1060 sample is in Quality mode, whereas that demo is in Ultra Quality mode.
It does look like AMD purposely left temporal data out of FSR to guarantee a higher level of compatibility across engines, and boost adoption that way.
There are 3 tiers of IQ/Performance: Performance, Quality and Ultra Quality.
The GTX1060 sample is in Quality mode, whereas that demo is in Ultra Quality mode.
That wasn't the problem with DLSS 1.0
DLSS 1.0 was state of the art DL-based temporal anti-aliasing, and because of this it was repurposed to work as a resolution enhancement tech, it has never been designed for that though.
So DLSS 1.0 did state of the art anti-aliasing, you can "emulate" such AA with DLSS 2.0 Performance + 4x DSR resolution, and then the result of DLSS 1.0 was upscaled from 1440p to 4K with another spatial upscaling neural network.
Of cause DLSS 1.0 was quite performance heavy (somewhere around 3-4 ms) and that's why brute forcing higher resolution was possible and quite obviously higher resolution 75-80% rendering with TAA (vs 66% with DLSS1.0) led to higher texture details with standard upscaling + sharpenning on top of the upscaled frame, because driver keeps up texture lods so that there is 1 pixel to 1 texel mapping on distant mip levels.
Even if DLSS 1.0 provided perfect 64x SSAA quality AA, there was no way for it to reconstruct higher levels of texture details in higher resolutions by making spatial upscaling, even the most powerfull spatial upscaling GANs would never infer REAL texture details from a single shot, that's why you need temporal accumulation in HIGH resolution space.
Spatial upscaling can only improve geometry edge quality (and DLSS 1.0 was really good at this) in case there is good AA in a game (it won't upscale aliased staircases well), it obviously won't help with texture details.
Hence, the whole AMD presentation looked like a pure marketing fight to me after the word "Spatial".
I can easily tell the difference with anything else(quite subtle differences between DLSS Balanced and DLSS Quality at 4k for instance) on a 1080p monitor on Youtube(using a 1440p+ setting), so I dont know why this should be different.
As for the 1440p 1060 Quality footage that looked bad - going by the performance scaling they're suggesting, FSR 'Quality' might be better equated to DLSS Performance or even Ultra Performance, though quality-wise, DLSS Performance is typically quite a bit better.
We definitely need to see quite a bit more to make any real firm judgements, though.
Panino Manino
Newcomer
cheapchips
Veteran
Not all games use TAA, even though a lot do. It makes it more universal as a technique? Maybe makes it simpler to add to a game?
FSR should definitely be more compatible. The question is it any good. Of course AMD knows that it will be subject to intense scrutiny so hopefully the fact they’re releasing it means they’re confident it will hold up quality wise.
Rootax
Veteran
Or they just think that they need to have something out, even if it's not as good as dlss at first. Like their RT, for now. I hope the quality is good. It would push nVidia to improve dlss even more I guess. And it would be good for everybody anyway.
