Nvidia DLSS 1 and 2 antialiasing discussion *spawn*

[manux]

Temporal part could be important on understanding what is going on. What is lost in resolution is gained partially back in more frames and less difference between frames?

 

[Mindtaker]

Just testing DLSS 2.0 on Control, amazing upgrade compared to 1.9

 

[MfA]

Yeah of course, we are now at a point where the user has to choose between two different rendering methods each with different set of strengths and weaknesses, with the caveat that one offers very large increase in performance over the other.

My point is that you can't point to what is sharpness as some perceived advantage when it's clearly partly aliasing. Aliasing only looks sharp, but it's not the same as sharp.

Staircasing is not super-sampling ground truth assuming you are sampling at the correct point in the pipeline. Although as I said it might be an artifact from using DSR to get training data.

 

[neckthrough]

Is the sharpening mostly limited to Control or is it just as prominent in other DLSS 2.0 implementations as well - MechWarrior, Deliver Us The Moon and Wolfenstein? If it's not present (or not as prominent) in the other games then perhaps it's just a stylistic choice. To those that dislike said choice (or even if you think it's intrinsic to the technique), why not provide some feedback to Remedy and NVIDIA? Help them all evolve this thing with some constructive criticism.

 

[Kaotik]

Is the sharpening mostly limited to Control or is it just as prominent in other DLSS 2.0 implementations as well - MechWarrior, Deliver Us The Moon and Wolfenstein? If it's not present (or not as prominent) in the other games then perhaps it's just a stylistic choice. To those that dislike said choice (or even if you think it's intrinsic to the technique), why not provide some feedback to Remedy and NVIDIA? Help them all evolve this thing with some constructive criticism.

If you go back few pages there's talk about similar issues with DLSS in Wolfenstein, so no, issues aren't limited to Control.

 

[pharma]

Is the sharpening mostly limited to Control or is it just as prominent in other DLSS 2.0 implementations as well - MechWarrior, Deliver Us The Moon and Wolfenstein? If it's not present (or not as prominent) in the other games then perhaps it's just a stylistic choice. To those that dislike said choice (or even if you think it's intrinsic to the technique), why not provide some feedback to Remedy and NVIDIA? Help them all evolve this thing with some constructive criticism.

For the most part those that dislike the sharpening improvement that DLSS 2.0 brings are very much in the minority. While more reviews are expected the next few days on the latest iteration of DLSS 2.0, those already out do not echo the dislike for DLSS sharpening heard from naysayers.

I actually like the sharpening effect and IMO makes the games more immersive and realistic, which think is one of the key DLSS 2.0 goals. Combined with the performance improvements it's pretty obvious it will be the preferred option for games and hardware supporting it. DLSS continues to evolve and get better and who knows what options might be offered in the future, but at least now there seems to be a pretty stable foundation to build upon.

 

[iroboto]

DLSS has come a long way. In a short period of time. On the same piece of hardware. It something that should be noted.
By design it can never be perfect; but everything in graphics is an approximation so far. People will embrace this as it continues to improve.

 

[pharma]

Testing Nvidia's DLSS 2.0: Higher Frame Rates for Free?
March 27, 2020

As you can see, the effect on performance of turning on DLSS 2.0 was substantial. At its most optimized, the 50 percent render image (Performance Mode) represents a 184 percent increase in frame rate over native 4K (54fps versus 19fps), while even the highest quality setting (66 percent) still grants a boost of 94 percent over the native resolution.

This trend continued in 1440p and 1080p results, all of which reflected an equally huge jump in speeds once the DLSS feature was turned on. Admittedly there is a bit of a bell curve where the percentage gains start to drop off in 1080p resolution, and we were told by Nvidia it was likely this would happen. At that point, the CPU is handling a lot more of the heavy lifting than the GPU would, which means the effectiveness of any tech contained in the GPU (the Tensor cores) drops accordingly.
...
My early impressions, though, are that Nvidia has created something special here, a nascent technology that could upend the value of graphics cards down the line. I'm a hard reviewer to please in this department; I've made my feelings about DLSS known on multiple occasions in more than a few of our video card reviews, and this is the first time I'm sincerely impressed with what the tech has to offer. And I'm not only impressed what it has to offer today, but also excited about the promise it offers for tomorrow.

https://www.pcmag.com/news/testing-nvidias-dlss-20-higher-frame-rates-for-free


CONTROL Gets DLSS 2.0, How Does It Look And Perform?
March 27, 2020

The RTX 2060 SUPER was used for 1080p because this is the class of card typically under the most scrutiny for its DXR performance and visuals as well as the 1080p resolution being one of the more difficult resolutions for the older Tensor Core based algorithm to work with. But we see a solid visual improvement over the older method and the performance is fine. We were told to expect the performance of the DLSS 2.0 to be a bit down from the original 1.X implementation in CONTROL and that was what we saw in the higher quality options. However, the visual improvement and consistency, especially in motion and around a metal chain link fencing the ever so slight performance penalty is worth it.

https://wccftech.com/control-gets-dlss-2-0-look-perform/


Edit: Additional DLSS 2.0 reviews
https://hothardware.com/reviews/investigating-nvidia-dlss-20-in-mechwarrior-5-and-control
https://www.overclock3d.net/reviews/software/dlss_2_0_with_control_-_nvidia_s_new_killer_feature/1
https://www.hardwareluxx.de/index.p...st-dlss-2-0-in-control-und-mechwarrior-5.html
https://www.pcgameshardware.de/Nvid...arrior-5-DLSS-Test-Control-DLSS-Test-1346257/

 

[dorf]

It seems NVIDIA is working on an adjustable DLSS sharpness setting. I noticed the command for it in Youngblood console earlier but changing it does nothing.

This is from NvRTX / UnrealEngine github repo (https://github.com/NvRTX/UnrealEngine access requires linking Epic account to Github):

https://imgur.com/raB9lJo

• Why doesn't r.NGX.DLSS.Sharpness doesn't seem to have any effect?
  • We are currently hard at work calibrating the user-adjustable sharpness setting to combine well with the internal sharpness value produced by DLSS's deep neural networks, in order to consistently deliver a high-quality output while still giving the user a significant level of flexibility over the amount of sharpening they want applied. It is currently available as a debug feature in non-production DLSS builds, but is disabled by default.

DLSS in Control is too sharp to my eye so hopefully this will setting will be exposed to players in the future. Youngblood luckily has it's own sharpness filter which works fine with DLSS but could be it isnt the optimal solution.

 

[dorf]

You can actually unlock the DLSS sharpness setting with a developer version of nvngx_dlss.dll found in the github repo. CTRL+ALT+F7 to unlock the console command in Youngblood and UE4 (r_NVIDIANgxSharpness and r.NGX.DLSS.Sharpness).

The difference between min (0) and max (1) is pretty subtle compared to the built-in sharpening in Youngblood.

Sadly this version of nvngx_dlss.dll seems to be incompatible with Control. The other games show a motion vector heatmap with CTRL+ALT+F12 but in Control the heatmap is dead and the image quality is poor too.

I accidentially stumbled upon these hotkeys with a hex editor as I was trying to remove an annoying "DLSS SDK" message that is burnt to the screen LOL.

 

[jlippo]

Like before, you can set custom DLSS resolutions in Control in the renderer.ini file. They work up to around 93% resolution scale after which the reconstruction will break down into a mess.

https://imgur.com/9j3hA3r

Yup, breaks when going too low.

Although at that point it's more of an curiosity than workable resolution anyway.
I was surprised how good 480x270 looked, 960x540 looked quite decent already. (On 1440p display.)

 

[Mindtaker]

GTC talk by Edward Liu from NVIDIA about DLSS 2.0, DLSS 1.0 comparisons.

https://developer.nvidia.com/gtc/2020/video/s22698

In this talk, Edward Liu from NVIDIA Applied Deep Learning Research delves into the latest research progress on Deep Learning Super Sampling (DLSS), which uses deep learning and the NVIDIA Tensor Cores to reconstruct super sampled frames in real-time. He discusses and demonstrates why scaling and image reconstruction for real-time rendering is an extremely challenging problem, and examines the cause of common artifacts produced by traditional up-sampling techniques. With this background, Edward then shows how the DLSS, deep learning based reconstruction algorithm, achieves better image quality with less input samples than traditional techniques. The talk concludes with a walkthrough of DLSS integration in Unreal Engine 4 and a look at DLSS in the acclaimed sci-fi adventure “Deliver Us The Moon” from KeokeN Interactive.



Digital Foundry analysis, DLSS 2.0, DLSS 1.9 comparisons.

Summary from ResetERA:

• DLSS 1.9 broke down under transparencies, resulting in ghosting and flickering
• higher resolution fixes subpixel breakup at the cost of higher cost (because higher res)
• 2.0 fixes subpixel detail issues like flickering
• movement also breaks 1.9
• ghosting trail still exists in 2.0 but is greatly reduced
• pseudo-random micro-detail (rock flecks, skin pores) is better preserved in 2.0
• text textures has higher contrast but is less legible than native res
• 2.0 can have high contrast edge breakup at times, but not really visible at regular zoom
• DLSS doesn't blur micro-detail in motion unlike TAA
• slight haloing with 2.0 (more visible at 800% magnification)
• SHARPENING IS TWEAKABLE IN THE SDK
• 1080p to 4K is 130% higher performance than native 4K in performance mode (4x scale) on 2080Ti
• 1440p to 4K is 67% better performance
• 2.0 cost more than 1.9, but in practice, it's marginally faster
• 1080p to 4K through DLSS has 11% lower performance than 1080p to 4K with regular upscaling and TAA
• on a 2060, that same test shows DLSS 15% lower than 1080p upscaled
• DLSS more expensive on lower end gpus
• 540p to 1080p DLSS resolves subpixel detail that a native 1080p image cannot
• Alan Wake ran at 540p on the 360
• halo artifact is more noticeable at lower resolution
• on a 2060, max everything, 720p to 1440p, runs in 40s in stressful environments (good for variable refresh rate monitors tho)
• using Alex's optimized settings from before, drops go as low as the mid 50s
• dropping to reconstructed 1080p, you'll stay above 60fps
• best image reconstruction solution so far, according to Alex

 

[orangpelupa]

Yoinks. This really made me regret getting a new gtx instead of a used rtx.

The DLSS 2.0 result is really good from DF tests, even on ridiculously low resolution.

 

[Scott_Arm]

Unless AMD offers a similar solution I can’t see myself skipping on Nvidia for my next upgrade. It’ll be interesting to see if dlss 2.0 takes off and is widely supported. Any idea if it’s difficult to integrate into games?

 

[iroboto]

Unless AMD offers a similar solution I can’t see myself skipping on Nvidia for my next upgrade. It’ll be interesting to see if dlss 2.0 takes off and is widely supported. Any idea if it’s difficult to integrate into games?

You have to create a separate render path that is completely native with no reconstruction or any form of anti aliasing applied. After that I think you just hook into their setup And you point to DLSS model file

 

[PSman1700]

best image reconstruction solution so far

I remember how much critique DLSS got from many, and see where we are now. It's probably just the beginning.

 

[Scott_Arm]

You have to create a separate render path that is completely native with no reconstruction or any form of anti aliasing applied. After that I think you just hook into their setup And you point to DLSS model file

Those conditions seem pretty strict. That won’t play nice with a bunch of games, notably Doom Eternal.

 

[Kaotik]

I remember how much critique DLSS got from many, and see where we are now. It's probably just the beginning.

I'm on of those critiquing DLSS and continue to do so until they bring DLSS 2X they promised on the launch already. After that I'll just critique DLSS "1X".
Yes, DLSS 2.0 does many things well, but for me it doesn't make up for the things it screws up in the process

 

[PSman1700]

After that I'll just critique DLSS "1X".
Yes, DLSS 2.0 does many things well, but for me it doesn't make up for the things it screws up in the process

The road to success wasn't smooth, or even a mess. What does that matter now?

 

[Kaotik]

The road to success wasn't smooth, or even a mess. What does that matter now?

Like I said: DLSS 2.0 does many things well, but for me it doesn't make up for the things it screws up in the process.
I'm not talking about the road there, I'm talking about the things DLSS 2.0 messes up, like causing ringing artifacts, overemphasized lines, apparently at least partly breaking DoF, missing details and specific objects turning into utter garbage