Credit to Wccftech, they've got a nice interview with Karthik Vaidyanathan - lots of interesting info IMHO:
https://wccftech.com/intel-xess-inte...-vaidyanathan/
cptwhite_uk (25-08-2021)
It all sounds very promising on paper - the benefits of DLSS2.0 (improved clarity over native res for fine details, decent low res scaling), while being GPU agnostic with support for older models, and generalised approach (so it doesn't need to be "trained" like DLSS, and operates more like FSR in that sense). I guess we'll just have to wait and see if this is all true, and what performance benefits it brings to the table.
Assuming it is true, even if the performance uptick isn't as impressive as FSR Ultra Quality, or DLSS Quality, I'd be happy to maintain image quality and gain a lesser performance boost.
It will be nice if all 3 approaches can have their own niche, so users can decide themselves where the sweetspot lies for them in terms of quality / performance trade off.
Last edited by cptwhite_uk; 25-08-2021 at 11:20 AM.
It still needs to be trained, but like DLSS 2, you can generalise models produced from the training (hence why you can swap DLSS 2 models between games and they still work for the most part). However it's certainly good that they've clearly gone for generalisation from the start and having the DP4a fallback is also good - seems to be a more 'benefit for more gamers' approach than NVidia's (unsurprisingly). Though of course, they're asking devs to start on the Intel specific acceleration first before letting them work on DP4a, so there's still going to be a bit of a hardware bias (I'd imaging targeting DP4a from the start would be better for everyone else rather than adapting something Intel specific later).
Me and a mate were comparing the images in the video,and in some instances the 1080p side look pretty bad,even for a 1080p images(the upscaled image was showing details "lost" in the 1080p one). I think we might need some more independent confirmation of image quality!
Where those details lost in a single screenshot or not present even in motion? If the former then that's exactly what temporal (super sample) methods are there for - you recreate the detail from accumulating previous frames, even if a single frame doesn't have the detail.
There are currently 1 users browsing this thread. (0 members and 1 guests)