By using this site, you agree to our Privacy Policy and our Terms of Use. Close
sc94597 said:
Mummelmann said:

This video has some interesting insights, and it's straight from the horse's mouth at that. Many issues that people bring up in this thread seem to be addressed, and it's worth a watch (hint; the detractors may have it right this time). 

I think one issue I have with this sort of video is that any discussion with any single Nvidia rep is going to be constrained on the reps end, especially when there is question-begging. 

Having said that, we do get useful information in this video, so I think it is very valuable for that alone. 

The conclusions are a bit hasty though. The framing of "only a 2D image and motion vectors" for example makes it seem like that isn't enough, when having access to the buffers is a big advantage DLSS 5 has over VLMs (which need to build heuristics in their features for object motion.) 

And yes, material properties and lighting need to be inferred, but DL models are pretty good at that, and have been for a while. 

What we've seen so far is a lot more temporally stable than early versions of DLSS 2, for example. 

The biggest issue is really the style shift issue. That will have to be fixed through artists, rendering engineers, etc working iteratively. It might also mean that they use neural shaders (which do directly access the material values) in their render pipelines when DLSS 5 is applied, to get that fined-tune control.

Even if they fix these issues it won't detract from the the fact that AI will kill us all one way or another.