WEBVTT

00:00:00.000 --> 00:00:06.480
AI upscaling is all the rage the most widely used being NVIDIA's deep learning

00:00:04.380 --> 00:00:11.639
super sampling I mean if you can keep good visuals with better performance why

00:00:09.360 --> 00:00:17.039
wouldn't you there's just one problem with dlss every game's implementation is

00:00:14.880 --> 00:00:22.020
different and they almost never get updated with the newest goodies later on

00:00:19.020 --> 00:00:25.080
if only there was some way to take the

00:00:22.020 --> 00:00:28.859
latest dlss and drop it into an older

00:00:25.080 --> 00:00:30.660
game oh wait there is and we're going to

00:00:28.859 --> 00:00:36.180
go over how to do it and what you can expect out of it like you can expect

00:00:32.840 --> 00:00:38.100
dang it earn out I got mad and hurt

00:00:36.180 --> 00:00:41.879
myself thanks to linode for sponsoring this video the note is a powerful

00:00:39.899 --> 00:00:45.719
affordable easy to use Linux based cloud computing service that is one click apps

00:00:43.739 --> 00:00:51.800
for games like Minecraft and valheim sign up today at linode.com Linus to get

00:00:48.360 --> 00:00:51.800
a free 100.60 day credit

00:00:58.100 --> 00:01:03.480
the concept is simple a game comes with

00:01:01.260 --> 00:01:06.720
a single file that determines its dlss version

00:01:04.320 --> 00:01:11.340
funny thing about files though you can replace them Tech PowerUp has a database

00:01:09.119 --> 00:01:15.360
of all the dlss revisions to date and programmer Brad Moore created a handy

00:01:13.260 --> 00:01:19.080
tool called dlss swapper that makes it dead simple to try different versions on

00:01:17.580 --> 00:01:25.259
your favorite games the question then becomes

00:01:22.320 --> 00:01:29.580
how far can we push it we start off by installing dlss swapper

00:01:27.479 --> 00:01:33.119
this would be straightforward but as a non-Windows store Universal app we need

00:01:31.680 --> 00:01:38.460
to import and Trust Brad Moore's certificate before Windows will let us install it so this is one reason why uwp

00:01:36.960 --> 00:01:42.600
apps aren't very well liked by the way but it's easy enough assuming we trust

00:01:40.439 --> 00:01:45.960
Brad here he's outlined all the steps you need to import it so follow the

00:01:44.340 --> 00:01:49.680
instructions and when you're done double click on the app to install it now we

00:01:48.119 --> 00:01:54.780
need to download the dlss files from Tech PowerUp directly which

00:01:51.840 --> 00:01:58.320
oh I'm sorry tech power up I really am for people playing along at home please

00:01:56.820 --> 00:02:02.100
try to spread out your downloads across multiple servers if you can and I'd

00:02:00.479 --> 00:02:05.759
suggest only getting the latest and oldest versions of each major revision

00:02:03.840 --> 00:02:09.239
to start off with you can always come back and grab others later once they're

00:02:07.740 --> 00:02:13.319
downloaded go ahead and click extract then optionally delete the downloaded

00:02:11.280 --> 00:02:16.680
files once that's done the software is pretty early though so there's no

00:02:14.640 --> 00:02:19.500
feedback when it's done extracting but if you want to be sure that they're

00:02:17.760 --> 00:02:23.459
there you can find the files inside your documents folder with that done we can

00:02:21.480 --> 00:02:28.200
finally do the swap well you can do this manually for any game with dlss dlss

00:02:26.280 --> 00:02:33.300
swapper only supports Steam games right now unfortunately there's also no filter

00:02:30.540 --> 00:02:36.840
for non-dlss games yet but you can tell when a game does support dlss by the

00:02:35.520 --> 00:02:40.620
version number on the bottom right corner this is the dlss revision that

00:02:39.060 --> 00:02:45.780
it's currently using and if you've never swapped the dlls before is the version

00:02:42.660 --> 00:02:48.000
that the game shipped with dlss 1.0 the

00:02:45.780 --> 00:02:52.800
first release rendered at a fixed 50 native resolution and it had no temporal

00:02:50.459 --> 00:02:57.480
component either so every frame got treated in a vacuum leading to

00:02:55.019 --> 00:03:01.620
shimmering artifacts and other errors that were often distracting and

00:02:58.920 --> 00:03:04.140
unpleasant 1.1 improved things slightly with a sharpening filter although

00:03:02.940 --> 00:03:08.400
there's not a whole lot more information about it and it wasn't around for long

00:03:06.239 --> 00:03:12.480
before 2.0 came out anyway that version removed the need for each game to be

00:03:10.080 --> 00:03:16.860
trained individually by NVIDIA and added the temporal element that makes steel SS

00:03:14.519 --> 00:03:20.700
so impressive to look at Now 2.0's debut game was control which notably shipped

00:03:18.720 --> 00:03:24.480
with version 1. that gave us the first side by side look at the two upscaling

00:03:22.680 --> 00:03:29.099
revisions and highlighted how important that motion Vector data was for

00:03:26.519 --> 00:03:32.580
improving image quality dmss 2.0 also added different scaling factors in its

00:03:30.840 --> 00:03:36.900
quality balanced and performance presets and dlss 2.1 added the ultra performance

00:03:35.519 --> 00:03:41.280
mode which renders at a third of the Native resolution not a lot is known

00:03:38.879 --> 00:03:44.640
about what 2.2 does yet but early reports indicate that tweaks have been

00:03:43.140 --> 00:03:49.080
made to reduce some of the or start effects we can see here that Doom

00:03:46.860 --> 00:03:55.140
Eternal Metro Exodus and War Thunder are all on dlss 2.1 while F1 2021 runs 2.2

00:03:53.159 --> 00:03:58.500
and of course shadow of the Tomb Raider runs 1.0

00:03:56.640 --> 00:04:02.220
clicking on a game will give you a list of all the dlss versions that you've

00:04:00.239 --> 00:04:06.180
downloaded and swapping is as simple as choosing one and clicking update as you

00:04:04.260 --> 00:04:10.439
might expect sticking to the same major revision of dlss is probably going to be

00:04:08.459 --> 00:04:13.680
the most stable option but if you want to be bolder you can easily reset it

00:04:12.480 --> 00:04:17.340
back to stock as long as you don't manually delete anything just like you

00:04:15.659 --> 00:04:21.959
could reset your water bottle to stock after getting our new spout lid but why

00:04:19.560 --> 00:04:25.800
would you ltdstewer.com half off if you already own one or buy it with a new

00:04:23.340 --> 00:04:29.880
bottle in testing I've found that as a general rule version 2.1 and 2.2 are

00:04:28.080 --> 00:04:33.900
often interchangeable but some games have a floor that's usually a few

00:04:32.040 --> 00:04:37.500
versions below what they ship with this makes sense because they'd probably have

00:04:36.180 --> 00:04:43.620
been initially developed with that version before shipping with an updated release I did run into one game in

00:04:41.820 --> 00:04:48.419
particular that didn't like any changes and that's War Thunder

00:04:45.300 --> 00:04:49.860
swapping the dll disables dlss entirely

00:04:48.419 --> 00:04:53.880
which may be a function of the anti-cheat software that Gaijin is using

00:04:51.600 --> 00:04:57.840
if that's true then be aware that other games with anti-cheat may or may not

00:04:55.620 --> 00:05:03.720
take kindly to tinkering with these core files either Doom Eternal which only

00:05:00.780 --> 00:05:09.660
recently got dlss and RTX support comes stock with 2.1.66 but runs fine with the

00:05:06.840 --> 00:05:14.840
latest revision down to as low as 2.1.55 in terms of visual Fidelity at 1080p

00:05:12.419 --> 00:05:19.320
Ultra performance the latest dll shows significantly less shimmering on this

00:05:16.979 --> 00:05:23.160
floor grading over stock which seems to corroborate the findings digital Foundry

00:05:20.820 --> 00:05:26.520
made in their dlss 2.0 First Look it's not likely that you'll be using Ultra

00:05:24.960 --> 00:05:32.699
performance at any resolution lower than 8K but this shows how far along dlss has

00:05:29.639 --> 00:05:35.759
come since its first outings F1 2021

00:05:32.699 --> 00:05:37.139
comes stock with version 2.2.9 which is

00:05:35.759 --> 00:05:41.340
pretty close to the latest revision available but still upgrading has a

00:05:39.600 --> 00:05:44.580
subtle effect on the amount of ghosting present on thin objects like the Pito

00:05:42.840 --> 00:05:49.080
tube at the front of the car this helps reduce often acting High motion Trails

00:05:46.740 --> 00:05:52.740
but in some scenes the car's wheels will still also ghost pretty badly and

00:05:51.360 --> 00:05:57.180
there's unfortunately not a whole lot that the update can do for that the

00:05:54.600 --> 00:06:02.340
reason for this is as best I can tell while the car is more or less stationary

00:06:00.000 --> 00:06:06.180
in the frame here the motion vectors for it are moving with it at high speed

00:06:04.080 --> 00:06:09.539
along the racetrack feeding irrelevant motion data to the algorithm this is

00:06:08.100 --> 00:06:13.500
visible to some extent around the potato tube at high speeds as well where it

00:06:11.639 --> 00:06:17.820
gets softer thanks to that ghosting even when you're not turning so does that

00:06:16.259 --> 00:06:23.639
mean that upgrading to a newer version is always going to look better maybe not

00:06:21.060 --> 00:06:26.880
in Metro Exodus we can again see ghosting thanks to that motion Vector of

00:06:25.620 --> 00:06:31.979
the train moving along the tracks interestingly it's worse with the stock

00:06:29.520 --> 00:06:36.539
and latest revisions of dlss while the older 2.1 revision seems to suffer less

00:06:34.259 --> 00:06:40.139
from it not only that but want the latest version provides better Clarity

00:06:37.919 --> 00:06:44.520
in the chain links in most cases the older version surprisingly often looks

00:06:42.419 --> 00:06:48.479
closer to Native rendering and I struggle to explain why perhaps the

00:06:46.800 --> 00:06:53.400
older version made less aggressive use of motion vectors than later 2.1 and 2.2

00:06:51.180 --> 00:06:56.819
releases this doesn't really line up with digital foundry's findings with a

00:06:54.960 --> 00:07:00.600
bright gun sight in the dark sewers so contrast and perhaps even color

00:06:58.800 --> 00:07:07.319
detection may make up the bulk of the tweaks since 2.1 but what about dlss 1.0

00:07:04.319 --> 00:07:09.240
well you obviously cannot use a

00:07:07.319 --> 00:07:12.900
different major revision here the game just doesn't launch in every game we

00:07:10.800 --> 00:07:15.720
tested however there were a few different revisions for each of the

00:07:14.220 --> 00:07:20.819
early titles that came out like Final Fantasy XV the debut title that shipped

00:07:18.180 --> 00:07:25.440
with revision 0.9 fascinatingly if we upgrade that to the final 1.0 variant

00:07:23.639 --> 00:07:29.039
released for shadow of the Tomb Raider which is I think Point 17 were treated

00:07:27.539 --> 00:07:33.539
to a slight image quality improvement it's pretty minor but the studs in

00:07:31.020 --> 00:07:37.319
Cindy's had here have a kind of smear to them with the stock dlss while the

00:07:35.580 --> 00:07:41.099
updated version shows them more correctly as round there's also a bit

00:07:39.419 --> 00:07:45.180
more contrast around her shoulder pocket and a bit more clarity to her jacket's

00:07:42.780 --> 00:07:48.120
patch you can much more clearly make out the Royal Capital lettering on the

00:07:46.740 --> 00:07:52.080
license plate too there's a lot of admittedly subtle

00:07:49.979 --> 00:07:56.280
differences and the Very earliest release shows even rougher upscaling so

00:07:54.060 --> 00:08:00.180
it's just a fascinating look to see how little changes to either the algorithm

00:07:58.620 --> 00:08:05.280
or the data set have improved things over time because remember these old

00:08:02.699 --> 00:08:12.599
revisions they were trained by NVIDIA on a game by game basis now you might be

00:08:09.060 --> 00:08:14.580
wondering if this easy to upgrade dlss

00:08:12.599 --> 00:08:19.139
why did Brad have to program anything well it seems to be that NVIDIA might

00:08:16.800 --> 00:08:23.580
want to look at opening up a little bit these off-label upgrades for dlss are

00:08:21.660 --> 00:08:27.720
fascinating and potentially can improve your gaming experience AMD's Fidelity FX

00:08:25.979 --> 00:08:31.620
super resolution may not be an AI upscaler but it's been Making Waves

00:08:29.580 --> 00:08:36.180
among enthusiasts and big projects alike thanks to its open architecture not only

00:08:33.539 --> 00:08:41.760
that but Intel's also open AI upscaler xcss hopes to compete with dlss's Walled

00:08:38.880 --> 00:08:46.860
Garden directly whenever we finally get a taste of Team blue gpus that's a while

00:08:44.219 --> 00:08:51.000
off yet but get subscribed because it'll be one of the biggest events in PC

00:08:48.480 --> 00:08:54.300
gaming Hardware in a long long time and you're not going to want to miss that I

00:08:52.980 --> 00:08:58.380
have to wonder where it all goes from here though we've got the scaling itself

00:08:56.580 --> 00:09:03.480
we've got the temporal element with multi-frame data we've got motion

00:09:00.480 --> 00:09:06.420
vectors to inform future frames what's

00:09:03.480 --> 00:09:11.160
next will the scalar take data about the viewport scene geometry

00:09:08.760 --> 00:09:15.660
I mean from there the question becomes how long before AI upscaling in effect

00:09:13.680 --> 00:09:20.519
is doing the bulk of the rendering work with so little input from traditional

00:09:17.940 --> 00:09:24.660
rendering that it's basically obsolete those are questions we can't answer

00:09:22.260 --> 00:09:28.980
today we also can't answer how swapping dlss might affect your favorite game

00:09:26.399 --> 00:09:32.220
there's just far too much to dive into for one video

00:09:30.180 --> 00:09:36.120
so put on your scientist goggles and maybe do a little bit of testing of your

00:09:33.540 --> 00:09:39.899
own it's super simple to do this video is brought to you by manscaped manscaped

00:09:37.920 --> 00:09:44.459
provides an all-in-one grooming kit that has you covered from head to toe their

00:09:42.420 --> 00:09:48.600
performance package 4.0 features their awesome lawnmower 4.0 waterproof body

00:09:46.620 --> 00:09:52.800
trimmer their weed whacker ear and nose hair trimmer plus a whole lot of other

00:09:50.760 --> 00:09:57.720
goodies and for a limited time you get all of this plus two free gifts the shed

00:09:55.500 --> 00:10:02.399
travel bag and a pair of manscapes anti-chafing boxer briefs so visit

00:09:59.760 --> 00:10:07.140
manscapes.com Tech or click the link down below to get 20 off and free

00:10:04.860 --> 00:10:13.320
shipping thanks for watching guys a small portion of this video's a-roll was

00:10:09.000 --> 00:10:15.180
legit AI upscaled to 4K from 560p did

00:10:13.320 --> 00:10:19.260
you notice if you're feeling confident take a look at our recent video on

00:10:16.980 --> 00:10:22.700
whether we could tell if dlss was on or off it was fun