Ok, then i guess DLSS includes some kind of temporal accumulation in itself and may also jitter the camera. The latter is something devs using it might know, even if the rest is proprietary and hidden from our curiosity...
Nvidia DLSS 1 and 2 antialiasing discussion *spawn*
- Thread starter DavidGraham
- Start date
-
- Tags
- nvidia
It learns what the image should be based on comparing the under-sized image to the perfect quality reference.
The multiple DLSS quality settings in these newer titles makes me dream of a dynamic DLSS resolution scaling. RT games would really benefit from this since the framerates tend to be quite low on average. Crashing from 70fps to 50fps feels much worse than crashing from 140fps to 80fps imo.
Its more like it's been 'learned'.
To continually learn is a different type of model. Not explored here with respect to this feature.
The target reference images used to train DLSS models are much higher resolution than the native resolution of your monitor. So yes it is technically possible if the reconstruction process is smart enough.
Now it’s another matter whether Wolfenstein is actually using a trained model or the hardcoded upscaling algorithm that Nvidia also calls DLSS. Talk about snake oil.
Technically they use a super sampling algorithm for training images. So at perfect DLSS quality it would come close to rendering SSAA for that resolution. Not sure if that is technically much higher than native necessarily.
it is possible for people to prefer DLSS and it also not be accurate compared to native. Haha. Sort of the pros and cons of deep learning. It’s just going to draw what it thinks it should be.
I wouldn't be surprised to see DLSS 2X once Ampere becomes available.
64x SSAA certainly is much higher
Again, who knows what they’re actually doing.https://www.eurogamer.net/articles/digitalfoundry-2018-dlss-turing-tech-analysis
Wouldn't that just be rendering at the higher native res? What does the 2X represent? Being 2x better? Being 2x higher res? Being 2x faster? Being 2.0 ?
I think it's meant to be cheaper super-sampling.
Oh 64x. That’s the first time I’ve seen that.64x SSAA certainly is much higher
https://www.eurogamer.net/articles/digitalfoundry-2018-dlss-turing-tech-analysis
Yea that’s quite a bit higher; though it gets resampled back down to the 4K resolution as I understand it. I agree on the who knows bit since there are so many levers to pull here.
2x faster. The model stays static unless it’s updated; so better hardware would run it faster.
Where are you getting that information? The 2x in the original DLSS documents from Nvidia was referring to a higher quality mode.
Sorry I didn't realize 2X DLSS was an actual setting.
For some reason I thought he was referring to a future 2X DLSS that does not exist yet because Ampere hadn't arrived. As in using the existing model with Ampere would net a better result.
I think this topic has been brought up several times. I think a lot of people just felt it was hot air until -- well today.
I had originally estimated that a 2070 would hav been perfect as a baseline console $399 as long as it could do both RT, VRS and DLSS. And I think honestly, it seems to hold true.
The fact that we turned the notch to 11 is pretty confusing lol. I guess DLSS has some drawbacks in the time to deliver. I'm not sure to be honest.
NVIDIA needs to eat their words on DLSS before that can happen, since according to them DLSS can only be used in situations where it yields performance gains. And DLSS2X can never offer that.
(unless they've already forgotten that decision)
