NVIDIA DLSS 3.5 Ray Reconstruction

[ZoneDymo]

I don't know if some people are blind or what but DLSS looks way better than native in every case.

watch the GN video and you will see that that is just not the case.

 

[zigzag]

The smoke that covers the building in the second comparison isn't gone, the DLSS just killed it. In the same fashion it added cell-shading to the cocrete pipes or over-sharpened the grid shadow when it's supposed to be soft.

Smoke is still there, but it has probably evolved into another state between screenshots. What you see at the edges of he concrete pipes could be bevel that has become visible due to improved sampling. Same with the grid shadow, undersampling of rays in native image could have produced blurred result. Its hard to know for sure without the ability to inspect the scene.

 

[Arco]

On my end, DLSS is kinda okayish? It's better than FSR but not by much. Still worse than native 100%.

 

[mrnagant]

I don't think it is a standout victory for IQ. Native in many ways still looks better. Even in the same image, DLSS makes one section of the image look better but then native in another area of the same image looks better.

I think the IQ is more like a wash between the 2, at least in stills, but DLSS obviosly gives a lot more fps.

 

[zigzag]

Why does better than Native keep getting pedaled in any DLSS review/marketing materials? By definition you cannot create more from less information; textures will never be better than native, not to mention a slew of other artifacts introduced from upscaling techniques.

Generative AI can create more from less information. It makes made-up details by making best guesses based on the information it has been trained on. These fabricated details are not applicable for purposes like forensic examination, but for games these made-up details can improve the immersiveness of the graphics.

Look at example how much details a bigger AI model can create:

Sadly, models this big don't yet run in real-time.

 

[Lew Zealand]

You are talking about that one? It's still soft, (compare it to the fence shadow) just not as soft it's also from a moving object since the position is also different.
View attachment 314567View attachment 314568

Shadows is not the thing that I would criticize with DLSS 3.5, but rather how it's altering some texture when it's fixing the Path tracing. Better trees reflections, but the marble is not marble anymore. (That's DLSS VS DLSS 3.5 btw)
View attachment 314569

In the first picture when comparing DLSS Q + RR with Native, the textures on the right hand wall are blurred out on the DLSS Q + RR cap vs the Native cap. I get that shadows can change and should look better with fuzzier edges at longer distances, but this is a simple well-lit wall and it looks like the texture didn't properly load in, like the LOD is broken. Same as in the marble to flat stone conversion in your screencap.

This is a simple and easily noticeable reduction in image quality, like the kind of problems I see in Starfield not loading in the highest rez textures on 6 GB GPUs. Except it's happening on a 24 GB 4090 and I assume because of the somewhat faulty Ray Reconstruction technique. We'll see if this is fixable like DLSS 1 was.

Generative AI can create more from less information. It makes made-up details by making best guesses based on the information it has been trained on. These fabricated details are not applicable for purposes like forensic examination, but for games these made-up details can improve the immersiveness of the graphics.

True. It's all computer-generated anyway so why not just add in a heaping helping of more computer-generation.

 

[gffermari]

Everything’s good about DLSS but it’s kinda worrying when you see DLSS+Frame Gen+Ray Reconstruction+…+…whatever tech comes next that guesses the frames or fills the gaps etc…

It’s worrying when you realise that the only real frame is a picture at “low” resolution and everything else is best guess from the tech.

Anyway, DLSS and Ray reconstruction are nice and I like them. This frame gen is still something I can not take in although it’s decent in the one game that I tried it (Hogwarts Legacy).

 

[lepudruk]

537.34 as recommended by NVIDIA for this specific article

That's weird.. CDPR on the other hand, claims that the new drivers (537.42) are mandatory. Let me quote them here:

To make sure DLSS Ray Reconstruction works correctly on your PC, update your NVIDIA Game Ready Driver to version 537.42 at minimum.

Source: <News Update 2.0>

 

[DY69SX]

It only works with PT enabled for now.


Youl learn to ignore them in time. It's people who bought the wrong GPU and now they pretend they can tell the difference. If you ask them for a blind test they will disappear into the ether.

Thx I figure it out and with PT I get round +-7FPS and with RR I get round +-18FPS Hue Hue Hue

 

[RainingTacco]

*Crying in AMD*

 

[rv8000]

Generative AI can create more from less information. It makes made-up details by making best guesses based on the information it has been trained on. These fabricated details are not applicable for purposes like forensic examination, but for games these made-up details can improve the immersiveness of the graphics.

Look at example how much details a bigger AI model can create:

Sadly, models this big don't yet run in real-time.

The problem youre overlooking is an actual designer/artist had a specific intent in a video game. Its one thing if the game was AI generated in the first place, but when there are actual textures/geometry made by hand the AI can only guess what was there. Numerous reviews and screenshots of DLSS and other upscaling techniques continue to show smeared textures, worse textures, garbled frames/UI artifacts (specifically related to frame gen), ghosting, and so on.

That doesn’t even begin to touch on the actual time it would take to AI generate in real time, not static images, like what you linked. Not feasible, whatsoever.

 

[zigzag]

In the first picture when comparing DLSS Q + RR with Native, the textures on the right hand wall are blurred out on the DLSS Q + RR cap vs the Native cap. I get that shadows can change and should look better with fuzzier edges at longer distances, but this is a simple well-lit wall and it looks like the texture didn't properly load in, like the LOD is broken. Same as in the marble to flat stone conversion in your screencap.

This is a simple and easily noticeable reduction in image quality, like the kind of problems I see in Starfield not loading in the highest rez textures on 6 GB GPUs. Except it's happening on a 24 GB 4090 and I assume because of the somewhat faulty Ray Reconstruction technique. We'll see if this is fixable like DLSS 1 was.

Yeah I noticed it too. In the first picture the texture on the right wall looks like a lower resolution was loaded. Also floor at that wall also looks too different to be due to RR. It might be random placement of detail textures or a bug. I'm guessing it's the latter. Some sort of bug with multilayer materials? Same with marble.

Shading of the trees and foliage, on the other hand, seems much improved with RR. Especially in the last picture.

The problem youre overlooking is an actual designer/artist had a specific intent in a video game. Its one thing if the game was AI generated in the first place, but when there are actual textures/geometry made by hand the AI can only guess what was there. Numerous reviews and screenshots of DLSS and other upscaling techniques continue to show smeared textures, worse textures, garbled frames/UI artifacts (specifically related to frame gen), ghosting, and so on.

That doesn’t even begin to touch on the actual time it would take to AI generate in real time, not static images, like what you linked. Not feasible, whatsoever.

AI generated upscaling, when it works correctly, doesn't seem to interfere too much with what artists have intended. But I agree that it has some issues. Biggest (and hardest to fix) are probably artifacts created by temporal processing. Many bugs that appear in games are fixable (like lower LODs being loaded due to lower rendering resolution), but this tech is still new and constantly changing, so until it matures and stabilizes, bugs will inevitably keep cropping up.

 

[r9]

It triples the framerate and you guys nitpicking.

 

[Nostras]

*Crying in AMD*

Oh cmon, if you bought AMD you did it because you wanted maximum raster/$.
I'm watching this news and will definitely be taking it in consideration for my next GPU, but for now I will be using none of that tech.

According to W1zzard I'll be getting 98 fps on my preferred resolution which is WAY below what I want (165) which means I'll have to tone down some if not most of the settings.
No room for these funnies.
None of that DLSS or raytracing shenanigans.
Just native.

 

[THU31]

Hmm. I have to say that some things in these comparisons look worse with ray reconstruction.

Like the texture quality on the right wall in the first screenshot. And the entire floor in the last screenshot, so much detail is gone.

Is this another example of something more realistic looking worse?

 

[oxrufiioxo]

Hmm. I have to say that some things in these comparisons look worse with ray reconstruction.

Like the texture quality on the right wall in the first screenshot. And the entire floor in the last screenshot, so much detail is gone.

Is this another example of something more realistic looking worse?

It's best to watch videos with movement some things look way better, some about the same, some worse. It's no different than any other tradeoff. It definitely not a blanket solution that is always better but neither is running the game rasterized vs ultra RT vs Pathtracing. The reality though is for most they don't have the hardware to even run it acceptably. I've used it in the first 3 hours of the game and overall it is a large improvement over not using it and I will continue using it for another playthrough and with the expansion.

For a first generation tech that works on all RTX GPUs I am pretty impressed overall but just like Frame generation it needs work. I am kinda glad it's not perfect though because if it did always look better than standard RT I would no longer have AMD as an option period as a gpu choice as it is in one game I much rather use it than not use it.

 

[csendesmark]

another in depth comparison of the pros and cons of using it although the person doing the testing says he will be using it at home.

Very good in depth video, and it reminds me when I played HL with Ray Trace
It is cool, but have a lots of bugs, which ejecting me from the immersion...

 

[oxrufiioxo]

Very good in depth video, and it reminds me when I played HL with Ray Trace
It is cool, but have a lots of bugs, which ejecting me from the immersion...

So far in actual gameplay none of them are noticeable at least not more so than terrible screen space reflections or RT reflections/lighting being generally more grainy and taking a few frames to update.

Disclaimer though that is with a 4090 running around 100 fps with frame generation on a 65 inch oled if I was sitting 2 feet from a 1440p monitor it may or may not be more noticeable

Definitely not any reason to purchase a RTX 4080/4090 for but def a really nice bonus.

 

[sLowEnd]

It looks good.

 

[Braegnok]

Updated to Version 537.42 today, installed Cyberpunk 2.0 update.

The new settings are much better on 40 series cards. With new DLSS Q + RR I'm getting 27 FPS,..

I need to run my old settings with 3090 Ti.

 

[zorandesign]

What a terrible naming scheme. So, my 3090 doesn't get the wanted FG feature with DLSS 3.5. Why should I even be bothered with this "HUGE" announcement? Meh. AMD is nicer to Nvidia users than Nvidia itself.

 

[Nordic]

Poor 2080ti and 3070 getting 2 fps. They are within margin of error from 0 fps.

 

[Metroid]

"Better than native" -> Proceeds to show a plethora of comparisons where DLSS borks effects like smoke, shadows, DoF or bloom.

Why is the conclusion longer than the rest of the article anyway?

Something is strange if is better than native, that means native rendering is broken. Native has to be the best possible image.

 

[InVasMani]

I'd stick to native + reshade myself. Better than native 100% additional post processing.

 

[oxrufiioxo]

Something is strange if is better than native, that means native rendering is broken. Native has to be the best possible image.

The problem is almost all game engines now use TAA as an AA method and if not implemented well can have more issues than DLSS it's probably better to look at DLSS especially in the quality mode as an AA method somtimes I prefer it over TAA somtimes I don't. Better than Native is always going to be subjective I try to say I like the DLSS image better than I like the TAA image the only thing I universally like better is super sampling but even with my 4090 that isn't possible in modern games at 4k but that's how I run older games with DLDSR.