Is DLSS 3 different from these AI methods?
polycounter lvl 12
i have been watching a lot of footage from the DLSS 3 presentation. Everything looks incredible but at the same I wouldn’t be convinced that it was GoPro or camera phone footage.
Now if someone were to show me something like this
or this
I would have thought everything I have seen was taken with a real life camera. Is DLSS 3 providing path tracing without mountains of gigabytes of reference photos? Is that why certain scenes look more realistic?
Replies
DLSS3 is taking 2 rendered frames and generating a frame between them based off motion vectors. It's essential frame interpolation or optical flow.
So DLSS3 is not the same as these other two projects?
https://www.nvidia.com/en-us/geforce/news/dlss3-ai-powered-neural-graphics-innovations/
no
the first one is style-transfer in realtime(ish) -
the second is someone making a game that looks like body-cam footage
DLSS/DLSS2 do 'AI' super-sampling, DLSS3 adds a temporal dimension to that.
Yeah, DLSS is an image upscaler. It renders at a fraction of the resolution and then uses AI to intelligently upscale to the target resolution. DLSS3 is the latest version of this, making use of some of the latest features in the Nvidia 40** series cards (possibly 30** too, I can't recall). It's a performance optimization more than anything.