DLSS Momentum Continues: 50 Released and Upcoming DLSS 3 Games, Over 250 DLSS Games and Creative Apps Available Now

[brostradamus]

how is it irrelevant any multi-layered neural network will have multiple passes by design, there are multiple layers that need to be computed in sequence

 

[mrk]

Because it doesn't matter if the instructions are or are not reduced (we don't know the full details).

What matters is the end result and the fact that even RTX 20 cards benefit.

So yes, it has no relevance.

 

[Proto]

The purpose of a neural network is that it’s not sequential.

 

[brostradamus]

Because it doesn't matter if the instructions are or are not reduced (we don't know the full details).

What matters is the end result and the fact that even RTX 20 cards benefit.

So yes, it has no relevance.

i would say thats inconclusive, maybe those original methods just needed some more optimization baked in, you never know if they are just selling a gimmick

The purpose of a neural network is that it’s not sequential.

thats possible in some cases where we are looking at NP hard problems, where the neural net is used to guess a solution thats close enough and classical algos could have taken much longer (for a more accurate result), but a neural net is fundamentally multipass because the layers have to be computed sequentially, the preceding layer is a mandatory input for the current layer

 

[mrk]

Ray Reconstruction is the optimisation. That's the whole point, to replace multiple manually implemented denoisers that are inefficient, costly by nature and can't be optimised due to the nature of what it (RT denoising) is.

 

[Proto]

i would say thats inconclusive, maybe those original methods just needed some more optimization baked in, you never know if they are just selling a gimmick


thats possible in some cases where we are looking at NP hard problems, where the neural net is used to guess a solution thats close enough and classical algos could have taken much longer (for a more accurate result), but a neural net is fundamentally multipass because the layers have to be computed sequentially, the preceding layer is a mandatory input for the current layer

Ok, I think there’s a slight disconnect in context.

Things happening in real-time are not sequential, but we try to tell a story based on the tools available. This meant that sequential computing became important.

A neural network takes away the sequence that you have in a simple loop, and tries to allow it to follow its own path. Like all true science, this can only be observed.

It can never be 100% accurate no matter how you look at it.

 

[Nexus18]

Alex is doing his first video on lossless scaling frame gen (and first frame gen comparison of officially decoupled amds frame gen):

 

[mrk]

Wtf is going on there with lossless

 

[TNA]

Wtf is going on there with lossless

Looks like it is losing something

 

[TheRealDeal]

Looks like it is losing something

Yep lossless some how seems like the wrong name for it.

 

[NZXT30]

Wtf is going on there with lossless

I've used it a couple of times and its utter garbage.

I struggle to understand where the hype for it comes from.

 

[TheRealDeal]

The FSR 3.1 also seems to be of the lossless varient. Dlss wins this by a land slide. I do see what's going on though Lossless and 3.1 give you some Jigsaw pices to fill in the gaps so you can get your kid in to have some fun. Will need to see it in motion though as that's what really counts.

 

[mrk]

It makes sense considering Nvidia has tonnes of training model data to feed DLSS FG whereas the others don't and are also relying on software methods only to generate frames.

 

[stellasuppa]

Looks like it is losing something

semi lossless

 

[Dicehunter]

Alex is doing his first video on lossless scaling frame gen (and first frame gen comparison of officially decoupled amds frame gen):

That FSR 3.1 image looks a little bit better on the sword than the DLSS version, That's quite impressive. Obviously Nvidia will have the edge in a lot of scenarios, No pun intended, Due to full on hardware acceleration but AMD's is quite impressive considering it's hardware agnostic.

Just a shame earlier games that have FSR won't get updated with 3.1 as it is a big leap in image quality over FSR2.

 

[mrk]

Are you looking at it right? The FSR one has more temporal image break-up, look along all of the edges of the sword, there is breakage and fizzle which is common with FSR when interrogating at a closer zoom.

From that example the DLSS FG version is arguably the cleanest/most stable output.

 

[Dicehunter]

Ahh yeah so it does, I was initially looking at the blade.

 

[TheRealDeal]

Are you looking at it right? The FSR one has more temporal image break-up, look along all of the edges of the sword, there is breakage and fizzle which is common with FSR when interrogating at a closer zoom.

From that example the DLSS FG version is arguably the cleanest/most stable output.

Even if AMD did in this image there is no top part to the sword either. Proper shambles but in motion would we see it no so sure. Terrible if it's visible while playing. The only acceptible frame there is DLSS. Glad i don't use any method but if i had to i want AMD DLSS hahaha.

 

[CAT-THE-FIFTH]

That is as lossless as the Tories Election Promises.

 

[Grim5]

Are you looking at it right? The FSR one has more temporal image break-up, look along all of the edges of the sword, there is breakage and fizzle which is common with FSR when interrogating at a closer zoom.

From that example the DLSS FG version is arguably the cleanest/most stable output.

Just look at the hands

There is a massive weird pocket of nothingness around the hands on the FSR image






In this comparison the quickest way to compare them is to look at the leafs, because the leaves are small background objects and moving, they are susceptible to temporal instability and the better solution will be able to display more leafs