DLDSR look sort of bad to anyone else or is it my configuration?

Just a hailmary question, did you capture the shots with Shadowplay?

 

Weirdly right now only 1.78x and 2.25x are available to my knowledge. Digital Foundry said this in their review of it iirc too.

 

Nope with windows game bar screen shot. I know that one attached image alone isn’t very useful but the full images are quite large and thus couldn’t be directly attached or put on Imgur.

 

Thanks I wasn’t sure what site would allow the full images. I’ll try that when I get back to my PC and could attach more full images. Good idea!

 

Then here's a shocker for you, nothing other than Shadowplay can capture DLDSR at its actual form (how you actually see on your screen). Only Shadowplay can.

Capture a screenshot with DLDSR at %0 smoothness and %100 smoothness with game bar, lightshot or Steam and compare them. You will see that they look similar, no difference at all.

Nothing besides Shadowplay can capture the special filtered look of DLDSR/DSR... There's some NVIDIA special stuff going on there that other software cannot see / capture.

I just wish NVIDIA had simple, nice looking scaling like AMD's VSR.

 

VSR is also softer than Nvidia 4x DSR 0% smoothness and thus I'd never use it.

Even Serious Sam 3 already had some sharp-tuned bicubic scaler at its disposal. Tweak weighting a bit to look less oversharpened or apply some Gaussian blur before the scaling takes place and it should look just fine. It's really ridiculous that we can't have that in driver.

 

DLDSR is great! I'd only use the Smoothness above 33% if there are no anti-aliasing options present in the game. 33% for x1.78, x2.25. At x4.00 (like others have said before me) 0%. Anywhere between x2.25 and x4.00 maybe 17%? You'll have to try it out and see for yourself.

 

VSR does not compare to DLDSR.

 

I meant how to manually config DLDSR to get beyond 4.00x. It's true the maximum figure is 2.25x, but I've seen some people played games at 10240 x 5000, 15360 x 8640, and above via DSR using just standard 4K monitor. Which makes me believed if DSR can be manually configed to such resolutions above 16K, DLDSR could be as well.

DLDSR > VSR > DSR, imo.
There were several comparison videos between VSR and standard DSR years ago, the majorities seemed to think VSR looks better.

 

I tried uploading the images I have to imgsli and the page kept timing out saying the files were still too large lol

It might work for just one or two at a time I suppose but going off of @oneoulker 's comment probably would be best just to retest at some point and use ShadowPlay so that the differences are better represented in the screenshots. I used Win11 Gamebar screenshot tool for what I have and I did only test one single game.

Ideally I'd love to see more outlets put out Digital Foundry style DLDSR vs DSR vs Native (in-game TAA on vs OFF, etc), but I find it rather odd (perhaps I'm just missing something) they're starting point smoothness value was 50% since in their own Witcher image 50% is still clearly a bit oversharpened vs ground truth. Maybe they were shifting more towards a sharper resolve since most games have a built in TAA which may soften the image somewhat? Unsure (they tested witcher 3 with in-game AA off).

I've rewatched that video recently and it's not really clear to me if Alex knew that smoothness with DLDSR works more like a reversed Nvidia control panel sharpen slider. In my own tests comparing 4x DSR 0% smoothness, native, and ground truth to my eye it looked like matching most closely to ground truth with DLDSR required a bit of a higher smoothness value though I tested this in Elden Ring only with in-game TAA enabled. I really do feel like there's just something "off putting" about the way DLDSR resolves at times. Or maybe it's Nvidia's sharpen being worse than CAS, I'm not really certain.

4x DSR with 0% smoothness looked more natural to my eye, but it still wrecked UI elements (which for whatever reason the borderless windowed mode -> win11/dx12 downscale does not do).

 

Just leaving a followup to this, I actually did retest this in another game (Hunt Showdown) and used Shadowplay this time, but at a glance I'm not confident its really capturing the DLDSR effect much better than the Gamebar screenshot (or, at least I can't tell the difference). Comparing screenshots in such a basic way (take 2 then zoom in on foliage, etc) seems less valuable perhaps then merely switching the game between native / dldsr / 4x dsr (though sometimes this requires lengthy reboots) and comparing that way then looking at it.

Interestingly in Hunt Showdown to my eyes at least I found a different amount of sharpness looked preferable so i'm starting to understand why DF recommended their 50% value as a ballpark starting point (I have not tested the witcher 3 with no AA which is what they did). If we're purely to go off of the DF image I attached before from their video then in that title it looks like 55% (or perhaps 60) without in-game AA would provide the closest match to ground truth/native sharpness. Antialiasing may throw that off somewhat, but even that's not gonna be consistent for all games since some games have really soft TAA resolve and other don't (or apply their own internal sharpen).

Kinda sucks we can't set custom sharpness values for DSR / DLDSR per profile and it also kinda is unfortunate we can't set a separate smoothness for basic DSR and DLDSR given smoothness seems to operate different between the two but ah well. Anyhow, all this testing is making me want to buy an actual 4K panel as I imagine a native 4K resolve will ultimately look much better than 4K DLDSR'd downwards to 1440p.

 

Any of you tried combining DLDSR with DLSS? I got some really good results in Marvel's Avengers.

So I was playing the Avengers yesterday on my 4K Ultrawide (3840x1600p) monitor with DLSS on Quality and GPU utilization was about 50-60% while maintaining the 116 fps limit (120 Hz G-Sync). I figured I should try and improve image quality since I have a decent headroom. I prefer DLSS Quality to native TAA so here comes DLDSR.

Did some calculations first and DLSS Quality runs at 2560x1067p before upscaling to native. I needed something better than that, while not being to taxing on performance. Both DLDSR options give me resolution of 5000+ horizontal and 2000+ vertical pixels, at 2.25x that is 5760x2400p. This is what I chose as the base and applied DLSS Performance. On 5760x2400p DLSS Performance renders at 2880x1200p. It's like a new Ultra Quality DLSS level.

With this I feel image quality improved drastically while maintaining the same performance. GPU utilization increased only to 70-80% which leaves even more room for improvement. Personally I don't find it necessary. The clarity and details this brings out is already amazing. The only thing that could use some fine tuning I feel is the smoothness level, as the default value of 33% was a bit to strong I feel.

 

DLDSR was never meant to have the 33% as a default from what I gather -- rather that was the default for the old DSR Guassian approach and since both can be enabled at the same time and share a slider (which is weird since smoothness looks to function very differently for both) they just carried it across (a bad choice seems to me).

Going off of Digital Foundry's tests matching DLDSR up with 4x DSR 0% smooth "Ground Truth" required setting smoothness to 50-60% percent (ingame AA off, not sure if they were using ingame sharpen or not but I assume it was off too). The image alex shared actually shows 50% with no in-game AA applied to still be a touch oversharpened vs ground truth. This is just in one game of course so this will vary as some games have very soft TAA implementations while others seem to apply heavy old/bad sharpen on top of their games automatically (Crysis 3 for example has a really bad old "dumb" sharpen algorithm on top of its SMAA 2TX and I assume Hunt Showdown too). We really should be able to tweak smoothness per profile, but in genereal terms 33% is way way oversharpened vs ground truth -- at least in my tests.

To clarify that, I mean "dumb" sharpen algorithms in the sense that they're not the CAS or modern algorithms which are much better. I'm not suggesting that sharpening a mild amount (especially after the image has been softened somewhat by aggressive TAA or something) is a bad thing altogether (though oversharpening can look mighty awful to my eye at least). I'm a rather big fan of CAS when used sparingly in conjunction with TAA (and ideally when both are built into the game so that UI elements don't get mucked up).

 

Since 4.00x DLDSR with 0% Smoothness would shame many Downsampling techniques, I wrote a request to Nvidia that we need 4x DLDSR. I doubt they'll reply to a suggestion though.

I'd encourage OP @BlindBison and others to write to them the features we think we need too.

 

It seems very weird that values above 2.25x are not available for DLDSR, yeah.

 

I use smoothness at 20-22% (atm 21%), and it seems to be the best by downsampled desktop - text, well I have to downsample desktop first to have a working ingame DSR by some games, or by old games to have 144Hz instead of 60Hz.

 

Is that smoothness with DSR or DLDSR? If it’s DSR then with in game AA 20-25% struck me as pretty reasonable in my old tests. For DLDSR 20-25% looked nutso oversharpened to my eye but maybe it helps with making UI text not get softened into smudges?