Death Stranding (PC) supports DLSS 2.0, allowing 4K at 60+ FPS on any RTX GPU

[RejZoR]

50 minutes ago, D13H4RD said:

It's not exactly making it from nothing as the model is trained using ultra high resolution samples, so it has an approximate idea of how to upscale the scene. 

 

And there actually are implementation of artificial intelligence-based upscalers like the aforementioned Topaz Gigapixel but also Photoshop's Preserve Details 2.0 upscaling algorithm and even waifu2x, which was specifically trained to upscale pics of anime styled images to a certain point. 

That is making shit from nothing. There is no way a prediction model can know what should be somewhere on the output side out of 700000000000000000 trillion possibilities of scene positioning within entire game. Or how something should look like, but doesn't and then it upscales it. Comparing that to upscaling of what's essentially a sequence of 2D images is just incomparable. I've done image and also video upscaling using using advanced algorithms and while they deliver pretty impressive results, they can only go so far.

 

Few examples of advanced upscaling I've done years ago:

https://rejzor.wordpress.com/2015/05/26/system-shock-2-high-definition-cutscenes/

 

And we're talking 2D images. Games are 3D dimensional data that's not recorded ahead of time but done in realtime for every tiny move player does anywhere within that 3D space. Saying you can train something to do that is just bullshit. NVIDIA is doing something far more simplistic behind the scene and then blowing it out of proportions by calling it "Ai whatever". Because if they were really doing some Ai godwork on the output, we'd have to run supercomputers for them to do these kind of approximations. And even then it doesn't make sense. Training the DLSS for a particular game however for it to render details in full and cut down the rest isn't some magic, it's just a matter of knowing how depth buffer for that title works and adjust the way how world in game is perceived by DLSS and make sure it renders as many things in full detail as possible while leaving unimportant stuff out to save cycles and increase performance. And they use tensor cores to do that kind of filtering. The training is for depth interaction and where details can be cut and not to train it how to magically bring details from low resolution. Because a racing game will have a different out of focus zone than FPS game or a 3rd person adventure and they can optimize titles for that.

 

AMD is doing a rather similar but more basic trickery with that dynamic resolution thing that's based on mouse movement. And I know they've improved it a bit to retain detail in static elements like GUI while scaling the rest so dynamic resolution based on mouse movement retains crosshair and HUD detail but downscales the rest. I'm willing to bet NVIDIA does something similar with DLSS, but in a more advanced way. And present it as Ai something because that's the marketing speak of today. Even if you're selling shovels, you're selling Ai shovels today.

[porina]

5 minutes ago, RejZoR said:

Because if they were really doing some Ai godwork on the output, we'd have to run supercomputers for them to do these kind of approximations.

My understanding of "AI" used in this context isn't great, but I can at least offer the following. In essence there are two parts two it, the training, then the application of the training. The training of the AI is the hard part. That's the part where you need a lot of compute resource regardless if you call it a supercomputer or not. Take the data, generate the AI model. nvidia does that part for you. The application of the training is much easier. That's what enables it to be run locally on consumer hardware. In other words, writing a program is difficult, but once done, running it is relatively easy. In this case it isn't a conventionally written program, but done through machine learning.

[Letgomyleghoe]

9 hours ago, RejZoR said:

If they can pull just what RTGI shader for ReShade can do that would be insane. Coz even though it's just screen space ambient occlusion and lighting, it creates such amazing effect on top of nearly any game. Games from 2009 feel like they have better lighting, shadowing and depth with it than current new games.

IK i was playing some cod mw3 (one of the last cod games worth playing) and the lighting was absolutely beautiful , you could see the shadows as leaves moved across the ground, bodys fell, the waves reflected, Fires creating a glow on some of the night maps. All that time I was thinking “holy fuck the only game rtx really makes sense is Minecraft”

[grayperview]

Nice that the game supports this, the game look poopy though (not graphically),

[Bravo1cc]

I'm so happy about this, one DLSS 2.0 is great, two Death Stranding is an amazing game, so happy its been ported well. It was great on PS4 but the frame times were inconsistent and resolution was okay. Roll on release day

[Vishera]

Well,My GPU supports RTX but doesn't support DLSS,so i am not that interested in it.

It's a cool technology though.

[RejZoR]

3 hours ago, Vishera said:

Well,My GPU supports RTX but doesn't support DLSS,so i am not that interested in it.

It's a cool technology though.

And what GPU supports RTX but not DLSS ? AMD doesn't have one yet and NVIDIA's ones should all support both or none.

[Vishera]

6 hours ago, RejZoR said:

And what GPU supports RTX but not DLSS ? AMD doesn't have one yet and NVIDIA's ones should all support both or none.

I have a GTX 1660 it supports RTX,I even played Control with RTX ON,DLSS is grayed out:

On 8/28/2019 at 5:26 PM, Vishera said:

I have results for 1080p

 

1080p,Max settings - RTX OFF:

 

1080p,Max settings - RTX MEDIUM:

 

1080p,Max settings - RTX HIGH:

 

Sorry for the picture quality (compression)

 

[porina]

10 hours ago, Vishera said:

I have a GTX 1660 it supports RTX,I even played Control with RTX ON,DLSS is grayed out:

I had forgotten about that. There was a driver update that enabled "RTX" on some Pascal/Turing GTX cards. So it's a kinda software support given it doesn't have dedicated hardware support.

 

Nice to see performance doesn't totally suck in that title, even if it is rather "cinematic".

[RejZoR]

That doesn't mean GTX 1660 supports RTX (or more specifically DXR in hardware). I can enable it on my GTX 1080Ti too in those few games in existence, but that doesn't mean it works well or is actually supported in terms of being hardware accelerated, not just "it can work". As it can be seen in your screenshots with massive framerate drop. Which is what I get too, maybe a bit less since my card is more powerful as is, but since there is no HW acceleration for ray tracing, performance hit is still massive. Also a dead giveaway is the GTX in your card's name. Just like on mine...

[Loote]

On 7/2/2020 at 4:44 PM, RejZoR said:

(...) Even if you're selling shovels, you're selling Ai shovels today.

Have you seen the effects waifu2x provides? Some images I could upscale 16x and the viewer would be no wiser that it was an upscale, of course on others the effects were no better than classic upscaling.

Just the way you upscaled those system shock videos, Nvidia can 'train' DLSS to use just the right methods based on comparing upscaled version with the same frame rendered at full resolution. It won't create detail out of thin air, but may create the best upscaled version for the chosen scene and with the resources of Nvidia it's guess should be solid. I suspect there will be many cases where DLSS would add something that isn't there or just make it look worse, but on the average it should be the best algorithm, if not now, then soon.

[illegalwater]

Here's a video from Digital Foundry covering DLSS and FidelityFX

tldw: DLSS looks better and runs significantly faster than native res TAA, also it totally destroys FidelityFX in visual quality