Nvidia DLSS antialiasing discussion *spawn*

it used to be pretty universally 1440p from 4K in 1.0. BUt according to wolfenstein, it is 1080p, 1296p, and 1440p for 2.0.

 

With DLSS 2.0 using it with 4K and performance preset the native resolution is 1080p:

https://www.nvidia.com/en-us/geforce/news/nvidia-dlss-2-0-a-big-leap-in-ai-rendering/

Dont know if performance is always "4X super resolution".

 

There's always the "DLSS-tax" aka time the tensor units need to make their calculations compared to naive scale

 

With Balanced mode enabled the value Youngblood console shows for rs_forceFractionY and rs_forceFractionX is just shy of 0.58 which would mean 1252p.

 

yeah, then that

 

Thanks. It seems the patch is now available for download.

https://www.dsogaming.com/news/cont...nload-adds-dlss-2-0-support-full-patch-notes/

 

Cool, the question is whether those same ratios hold at DSR factors too.

On a 1080p monitor with DSR 4x and DLSS 2.0 performance does it render at 1080p, upscale to 4K using DLSS and then downsample to 1080p?

 

Do we have performance data on this?

 

DSR renders at higher resolution and downsamples. I'm not sure when the driver is downsampling though, is it some sort of post-process? Not sure which would be applied first (DSR or DLSS)

 

Just don't go comparing the shots they posted and instead look at the originals.
I've only looked one so far (the last pair, receptionist or something) and at least in that particular one there's obvious ringing artifacts and some really odd ones, like the row of darker dots in one of the hinges and high contrast "raster dots" on the posters alarm clock, or the incomprehensible "rules" text in one of the posters

 

I'm not sure why they come to that conclusion. Looking at the first couple of their screenshots in 4k, the native is much better.

 

Their screenshots left a lot to be desired. Definitely will need to wait for a better DLSS 2.0 review.

 

Are you sure you didn't misread the labels? The placement of the labels between shots makes it harder to determine whether the label is referring to the shot above or below it. For every pair, the top shot seems to be with DLSS on and the bottom one DLSS off (native). For the first pair:
DLSS: https://www.hardwaretimes.com/wp-co...Control-Screenshot-2020.03.25-19.28.04.46.png
Native: https://www.hardwaretimes.com/wp-co...Control-Screenshot-2020.03.25-19.28.28.30.png

 

I think they learned a bit from AMD ... lots of people like a bit of oversharpening.

 

According to NV there are no sharpening and this result is what NN outputs.

 

That's not saying much, if you train it on super-sampled frames with a bit of sharpening the NN will sharpen. Or maybe they just add a bit too much LOD bias for my taste?

Or maybe it's the game ... it's possible that because of non-linear colour operations super-sampled frames actually make the textures look aliased. Can't blame DLSS for that. Comparisons should add super-sampled/DSR frames as a groundtruth for comparison.