Caporegime
- Joined
- 9 Nov 2009
- Posts
- 25,901
- Location
- Planet Earth
Nvidia responded directly to Daniel Owen's video:
www.reddit.com
He asked some very specific questions and goes through the responses.
So it does look like in order to apply the better lighting models,it is recreating the image with the better lighting baked in. But,it appears it is doing this on a per 2D frame,with some idea of motion vectors.
But dependent on what models are being used currently,what data sets they have been trained on(Instagram?) or the amount of data which is being used,it is making mistakes.
This sounds a bit like the DLSS1 scenario - will hold off final judgement for the full release,but it wouldn't surprise me if this undergoes a bigger overhaul under the hood closer to the date.
Reddit - The heart of the internet
He asked some very specific questions and goes through the responses.
So it does look like in order to apply the better lighting models,it is recreating the image with the better lighting baked in. But,it appears it is doing this on a per 2D frame,with some idea of motion vectors.
But dependent on what models are being used currently,what data sets they have been trained on(Instagram?) or the amount of data which is being used,it is making mistakes.
This sounds a bit like the DLSS1 scenario - will hold off final judgement for the full release,but it wouldn't surprise me if this undergoes a bigger overhaul under the hood closer to the date.
Last edited:




But there's also a difference between improving details and replacing characters completely with some new ultra pretty ones, that's seems very popular. I'm sure a lot of people would love to have sliders added to dlss 5 to improve certain geometry too - if you catch my drift
And yes, training data seems to be the biggest problem currently as that's what influences how it changes faces. All the big corporations working with AI just scrape social media (which contain a lot of fake AI generated profiles and photos these days!) and maybe buy some stock photos. Hence none of those look very realistic, as it's trained on very biased data.