Nvidia DLSS antialiasing discussion *spawn*

Yeah initial impression was that developers provide the training screenshots but the latest Q&A indicates that the "DLSS team", which I'm assuming is Nvidia, extracts screenshots from the game. It's all written by PR folks so it's hard to be sure exactly what the process is between the devs and NV.

 

So outside pre-scripted renders DLSS looks like FXAA, but it's super expensive and much more limited.

No wonder the support in games has been taking all these months. nVidia seems to be desperately trying to train those neural networks over and over to see if they can extract anything other than a dud, with no success so far.

Oh well... at least the tensor cores are apparently still useful for offline rendering.

 

Aren't they there for gaming as well for faster de-noising? I don't think anyone is using them yet but we don't exactly have a lot of RTX enhanced titles.

 

In gaming, the tensor cores are only being used for DLSS AFAIK.

 

its actually pretty fast I think, at least relatively speaking to the total development time of the title. The relatively very little labour as well comparatively to the labour costs of implementing another AA technique. The only thing missing is the quality which nvidia will be working on.

 

Detailed analysis here for the DLSS upscaler aka blurryfier
https://www.techspot.com/article/1794-nvidia-rtx-dlss-battlefield/
DLSS no match for TAA in BFV.
This starts to be looking like an expensive waste of silicon.

 

I agree 100% that DLSS looks terribly blurry right now and, in metro's case, nothing at all like 4K - but there are some positives as well that should be mentioned. It looks more temporally stable than TAA in both games and does not ghost on things like GPU particles or other alpha.

 

This is where using the learning model can help: they can use the pixel inputs rather than just the final colors to decide what to do.

 

In BF5 TAA doesnt affect reflections in windows. DLSS looks a lot better (less aliasing) and is more stable (less ghosting).

 

Better if someone prefers blurry images, but it seems i am the only person in the world who does.
Looking at the turret at given time in the video, DLSS has more crawilng to me, so i would not call it more stable.
Curious if NV manages to fix this and the sharpness as well... maybe there will be still some players left then.
Otherwise they better find another application for the silicon.

 

I wrote specific about the reflections in the windows.

 

Interestingly at least on top of surfaces where RT is used leaves and particles cause ghosting even on DLSS.

 

Maybe that's ghosting coming from RT denoising? I would not think upscaling can cause or prevent ghosting at all?

I did not spot ghosting myself, but i saw a whole tree disappear - looked like dithered transparency to me, and the sudden vanishing might be caused by sync issues of destruction in multiplayer. (The dithering can help to hide popping caused by such things and may be intended.)

 

RT denoising would be my guess.
Although I wonder how good job DLSS would do without temporal part.. it could be surprisingly effective.

 

With the patch DLSS looks more like the FF15 than BF5 implementation. A little bit to aggressive with the sharpness but the quality is so much better.
Here is a quick comparision between 4K/DLSS and 4K/0.6 shading rate (~FPS):
https://www.resetera.com/threads/metro-exodus-pc-performance-thread.99817/post-18158206

With Raytracing 4K is out of question with a RTX2080TI. But DLSS give a good improvement over using the shader rate or playing in 1440p (which is still much faster).

 

You mean 10% faster, right?

 

So did nVidia tweak the output of its last "training session", or did it just complete a new one? Do what extent can the new training session build on the old one, in regards areas that were covered the first time?

I ask because this improvement comes pretty quick on what we saw released not too long ago. If new "training sessions" can be configured, and then implemented, in under a week or so, then that could be promising news, if we use what's being reported on here as pertaining to the whole game.

nVidia only has to support a few games right now, so I wonder if their super computer that runs these training exercises is under expansion.

 

I'll try to find the link but read that they "update" existing trained models with new samples. I guess the same principle used in AI training and inferencing, and will likely be explained in greater detail when the developers & Nvidia share their experiences.

 

Warning, cooking analogy up ahead! lol So it might be like a stew. You serve it up, and it gets some criticism. Rather than throw the whole batch out, you spice it up, add in more broth so as to dilute its saltiness, and add in some pasta to give it texture.

P.S. Bonus Trivia: According to legend all hobo stew can trace its roots back to the first batch of Mulligan stew from hundreds of years ago. Every time hobos break camp everyone who is able takes a little of the always simmering stew along, and they use that as a starter in the next hobo encampment. Newcomers bring whatever they can, and toss it in the pot.

P.P.S. I am posting while hungry! I'll maybe come up with a better analogy after having eaten.

https://en.wikipedia.org/wiki/Mulligan_stew_(food)