Nvidia DLSS antialiasing discussion *spawn*

UPDATE: Miscommunication apparently, they were talking about FidelityFX (so the sharpening filter), not super resolution:

https://videocardz.com/newz/metro-exodus-will-not-support-amd-fidelityfx-super-resolution-fsr

 

ML is such a fuzzy term, at the moment it's basically equivalent to MLP ... but MLP is almost certainly not the best performing way to make a trained interpolation weight generator without "free" TFLOPs of matrix multipliers available.

 

I typically associate anything ML with not MLP. There's a lot of useful regressions, random forests etc, that I don't often see the need to move to MLP. You've really need to have 'one' of those types of problems.

 

Call of Duty: Warzone DLSS Benchmarked: 8K 60FPS on GeForce RTX 3090 | TweakTown
May 3, 2021

 

Nvidia has published a full list of DLSS games here:
https://www.nvidia.com/en-us/geforce/news/nvidia-dlss-games/
Unfortunately, this list is a mix of DLSS 1.0 and DLSS 2.0 games without any separation.

 

https://www.computerbase.de/2021-05/metro-exodus-pc-enhanced-edition-raytracing-benchmark-test/4/
The pros and cons of DLSS in Metro Exodus
DLSS 2.0 works much better in Metro Exodus than the original DLSS. And, unlike in the original version, it makes sense to use DLSS if the performance with the native resolution is not sufficient. Otherwise, Ultra HD and ray tracing cannot be played smoothly even on a GeForce RTX 3090. But you have to be aware that even with the highest setting you are sacrificing some image quality.

On the one hand, DLSS is on the right track in Metro Exodus with the successful image sharpness and the better drawn details, but the smearing, the more restless image display and the problematic ray tracing reflections are disadvantages that cannot be discussed away. This means that DLSS 2.0 in Metro Exodus is the only real emergency solution in the event of poor performance, but also only an emergency solution.

The test has shown that the DLSS setting "Quality" or "Balanced" is advisable in Ultra HD. For WQHD, “quality” should be targeted.

https://www.computerbase.de/2021-05...-raytracing-benchmark-test/5/#abschnitt_fazit
DLSS is the solution for 4K, but it is also fraught with problems
Alternatively, Nvidia's DLSS 2.0 can also be used on a GeForce RTX, which works much better in Metro Exodus than the first iteration and can be used sensibly. But there are limitations. DLSS leaves a good to strong impression in terms of image sharpness and the drawing of fine objects. The latter in particular is a great strength and there are a lot of fine objects in Metro Exodus. With native resolution, for example, branches like to disappear into nothing - DLSS does this much better.

However, DLSS also shows some weaknesses. The fine drawing of small objects is a strength, but they flicker slightly even in Ultra HD, which is not the case with native resolution. There, DLSS is usually superior to the native resolution, in Metro Exodus it is the other way around. In addition, the fine details in the game often tend to "smear", which looks ugly when in motion. The ray tracing reflections are also embellished, but the AI upsampling reacts precisely to the denoising problems of the reflections and exacerbates them. This means that DLSS 2.0 leaves a much better impression in Metro Exodus than DLSS 1.0. But it is not more than average. For smooth gaming in Ultra HD, there is still no way around DLSS without reducing the general image quality. Nvidia's intelligent upsampling in Metro Exodus has problems with this, but smooth gaming in Ultra HD is still the only way to get it. Here, AMD has to make improvements as soon as possible.

 

In the "open" world part of the game even native resolution has problem with flickering. Right from the beginning in Volga it is visible. So DLSS2.0 is not a "emergency" setting it is an alternative to native with better and worst aspects.

 

Or maybe, just maybe, it's possible that not everyone agrees with your exact point of view and do see it as worse than native in this (and other) cases, not as an equal or better alternative

 

Maybe i expect that something which is 40% slower delivers a noticeable better image quality.

 

The difference clearly is noticeable for ComputerBases review team, since they point out that DLSS should in this case be used only if you're not hitting sufficient performance without.

 

I would be disappointed if there weren't any "edge cases", as in the ComputerBase review.

 

Sure, when i compare images over images over images. But i play games. And the difference between native 4K and 4K@DLSS Quality isnt relevant outside of edge cases like more noise in few situations and a few cases of ghosting. On the other hand i have 60%+ more performance...

 

Since DLSS is an external dependency, it is solely up to Nvidia if they want to fix the bad interaction with Metro Exodus. It's somewhat of a surprise that there were still perceptible shortcomings given that DLSS has already been implemented twice in the game ...

 

There are lots of "edge cases" with TAA too, it just that nobody pays attention to them.
Shimmering on foliage, please. DLSS for comparison.
Wrong foliage and missed thin details (1440p).
TAA upsampling exhibits these flaws even further - what's about severe geometry edges flickering in motion with 70% TAAU in WD2 or shading aliasing for that matter?
How is anyone supposed to play with all those flaws?

That's not always the case. Neural Nets are trained on certain data with certain parameters. NN will not work if new data is out of learnt range.
There are 2 ways to fix this issue:
1) Retrain NN with the new data
2) Tweak your data so that it works with NN
In case of images this can be gamma for example.

 

It's an ill posed problem, lacking some essential data just to make it a drop in TAA replacement, training can only do so much.

 

https://www.nvidia.com/en-us/geforce/news/may-2021-rtx-dlss-game-update/

 

Woah, DLSS in VR

 

The links don't seem to be working.

 

Fixed. Hopefully links won't change this time.