DLSS in Control Tested

DLSS in Control Tested

Welcome to a bonus video for all of my observations
when testing DLSS in Control, and how it compares with when it’s disabled. Although Nvidia
would describe DLSS as an upscaler, I’d rather describe it as a mixture of anti-aliasing
and sharpening, and it certainly tackles the problem in some new and rather interesting
ways, which I’ll show in this video. It’s meant to use AI trained on high-res images
of the game to intelligently blur jagged edges, but I’ve heard it might not even use that
in this game. It doesn’t matter. I’ll judge it on how it performs in the game, using
a Geforce 2080 TI. Running Control, on absolute maximum settings,
at native 4K is… not a good idea. I got 20-25 FPS and it wasn’t a lot of fun. Dropping
to 1440p with DLSS increased it to a very playable 40-50 FPS which I felt was okay,
though disabling a few of the more problematic raytraced features certainly smoothed it further
and edged it closer to the magical 60 FPS mark.
And at 1080p, it remained well above 60 FPS at all times even with everything on. But
again, remember I’m using a 2080 TI here so you’d hope this would be the case.
But for this video, most of it will be rendered at below 1080p… and then zoomed in. Because
a lot of the changes DLSS makes are at the pixel level and this isn’t too visible when
viewing the screen as a whole. Especially what with Youtube coming along and being all
compressed and stuff. So just remember that what’s shown in this video might not be
quite so noticeable when actually playing the game at a proper resolution.
So how is DLSS in Control better than it used to be? I think the main reason is that, at
least now, it’s honest. It doesn’t just claim to be 4K while actually running at a
lower resolution. It shows you what resolution it’s running at, and lets you choose different
base resolutions. And it works at any screen resolution. In other words: while these improvements
still don’t give DLSS any real benefits over other methods, its implementation in
Control does remove a lot of its disadvantages! This made it easier for me to run about and
to see what DLSS actually does to the image. So you know: I don’t think you can run Control
without anti-aliasing. You can choose to use MSAA, DLSS, or without either it still looks
like it uses temporal. So in this video I’ll be comparing DLSS to that. Temporal is already
quite an effective form of anti-aliasing that’s widely used in modern games. To summarise:
DLSS typically runs 1-2 FPS slower, which depending on resolution can be somewhere between
2-5%. I could have sworn the difference was greater than that when I first tested it,
but I tried all kinds of resolution combinations and it remained, at most, just a frame or
two slower. DLSS was better at retaining detail on things when in motion while TAA would pixelate
the whole thing. But this came at a price- it introduces visual ‘bugs’- you can make
out halos around things in the foreground, which I think is it attempting to use data
from past frames and somewhat failing. I think you’ll agree that the plants in the foreground
look BETTER with DLSS. They retain their smooth edges and the definition on thin blades of
grass. It’s just that it’s at the expense of the background. DLSS appears to prioritise
things in the foreground, and it’s down to personal preference whether you think that’s
an improvement. The visual artefacts reminded me of the effect
you get when converting 30 FPS videos into 60. Obviously slowed down at a base resolution
of 720p is a worst-case scenario- if you use a higher base resolution like 1080p or 1440p,
and run the game at normal speeds then it’s nowhere near as noticeable.
Anyway, by spending endless hours, running about these same boring old corridors at different
settings and resolutions, I’ve grown familiar with DLSS and how it works. I can now identify
when it’s on, and wish to share with you what I’ve learned.
Comparing the first frame of a new viewpoint, where it has no past frames to help anti-alias
the image, reveals how DLSS works- when faced with a fresh, aliased image, it paints a low-resolution,
defocused waterpainting of it that blots out high contrast spots, then fills in the details
over the next dozen or so frames. The first frame is technically inferior to regular anti-aliasing
techniques, but I think to the human eye, it looks BETTER by minimising the number of
stray, high-contrast pixels. We don’t notice the water-paint effect on the first frame
as much as we would a high-contrast area that hadn’t been anti-aliased! I say it takes
a dozen or so frames for it to reach maximum quality, but within a fraction of that it’s
already pretty much there. Its defocused approach reduces the speckly
look of noisy surfaces, and while both methods suffer from ‘flicker’, I think DLSS can
end up using more individual frames than temporal does as its ‘flicker’ is slower. This
sometimes worked against it, producing a far more noticeable flickering effect, particularly
on this transparent glass- which unfortunately for DLSS is EVERYWHERE! AHAHAHAHAH!!!!
So it’s not a clear-cut victory for DLSS. Apart from in THIS example, it’s generally
a cautious type of anti-aliasing- it would rather blur, or even HIDE, high contrast edges
entirely. A perfect example of this is on the wires holding up the lights here! Temporal
shows a rather aliased result, while DLSS chickens out and effectively removes them
completely! I thought the edges of these pipes were interesting.
DLSS removed jaggies more consistently than Temporal could, whilst somehow managing it
with less blur. I think sharpening deserves some praise here- you can definitely see the
contrast being upped along edges in the DLSS example, almost as though Nvidia are borrowing
from AMD’s FidelityFX. Despite Temporal’s greater amounts of blur, you can still make
out the pixel-stepping along some edges. Not sure if this is from the anti-aliasing or
from the upscaling… but luckily, since I’m only comparing the end images… I can simply
say that DLSS does it better. There’s a graphics feature in Control that
really doesn’t play nicely with DLSS- and that’s the Ray Traced Indirect Diffused
Lighting in this game. In fact, it comes with a whole host of its own problems as well.
I guess it’s a more accurate simulation of ambient occlusion, but it’s not really
worth using because it adds smeary artefacts around anything moving. It tanks framerates,
and when DLSS is enabled, it makes some surfaces look horribly pixelated. The worst offender
being this wall here. DLSS appears just to upscale this effect without
making any effort to smooth the result, meaning you’re left with obvious low-resolution
squares across some of the walls in the game. I prefer TAA’s solution to it, which is
to blur it all back to 1999! Since it does tend to run 1-2 FPS slower than
TAA, and doesn’t always deliver superior results, I wouldn’t say that DLSS is a clear-cut
victory in Control, but it certainly deals with aliasing that Temporal might otherwise
miss and I do find something appealing about such a novel solution. If you’re anything
like me then, before this, you may have assumed that DLSS adds ‘detail’ to the upscale,
when in actual fact it just offers a relatively alias-free solution that can, if given enough
time, provide a low-blur, high quality result, with fewer upscale artefacts than with Temporal
alone. But when DLSS goes wrong, it goes VERY wrong.
So that’s a taste of the DLSS implementation in Control. See the main video for more information
on it!

Comments (56)

  1. Thanks Phillip, very cool!

  2. Wait, it’s not csgo? Oh wait nvm I’m on klicks not 3klicks. I’m just subbed to too many of your channels

  3. haha first dislike me funni snart

  4. 6 views 10 comments 24 likes. Seems legit.

  5. Phillip I beg of you do not stare into the darkness behind your eyelids

  6. Hey Phillip,

    i think you check the comments once a video has been uploaded so i wanted to use the opportunity of being this early, to simply compliment you on your work on youtube.
    You are producing great videos man 🙂

    Greetings from Germany

  7. DLSS in Control Tested

  8. How can machines know logic, it's fucking amazing.

  9. where can i listen to music that is in your video?

  10. was about to sleep since I'm horribly sick and need some rest. But wait. What's this? 2 new videos from daddy Philip? ?

  11. Ha you have the same GPU as me.
    What are you planning on saving for next gen if at all Phillip?

  12. This may be a dumb question, but what exactly is DLSS?

    Edit: Nevermind, found the other video 😉


  14. Is that ghosting from ray tracing? Or TAA? I would rather play at native 720p and no ray tracing than having that ghosting.
    Every time I make new project in Unreal Engine 4 I got to turn off TAA instantly because its completely horrid imo, due to the ghosting.

  15. The difference is barely noticeable if im being honest

  16. From what HWU found it's more like upscaling with some method of aa
    Not good

  17. RTX will be obsolete in 10 months.It will be irrelevant,but its a cool gimmick…at least for now.

  18. A 20 second unskippable ad appeared. TWENTY SECOND. UNSKIPPABLE. ADVERTISEMENT.

  19. I'm trying to watch a video about up scaling from 720p to 40 on a 720p display

  20. I really like these technical videos, especially because theyre short, but dense. I feel like ive learned something

  21. Nice looking game isn't it. The ray traced lighting looks very nice.

  22. what game is this?

  23. What about the Table physics though?

  24. TAA is fucking trash! it blurs images in motion, ALWAYS! it makes me sick, absolutely unplayable

  25. I'd rather have jaggies then forced anti-aliasing that makes everything blurry and cant be turned off in the settings menu or the ini file

  26. 90fps should be the standard now :c

  27. what game is this ?

    i guess im play csgo too much, make me know less about other games

  28. Gothic 2 Gold Edition

  29. Control looks so fun wtf

  30. Loving this vid on 144p

  31. And here I am with my 6GB of RAM and single core CPU…

    And Nvidia GT 620 …

    Ah well, at least I can still run Portal 2 a medium settings

  32. under 30 fps with a 2080 ti? :O What will i have with a 2060 on 1080p? xD

  33. He points out things that look blurry or jagged, but it all looks good to me.

  34. while you were talking about AA i was way more impressed with how detailed the destruction is in this game.

  35. another technology to pretend we can somehow do without subpixel information and leave consumers unsatisfied. just a pretense to another half a dozen generations of graphics cards where then they'll finally gradually get better at super resolution rendering to /actually/ solve the problem.

  36. The game just felt like a techo demo to me tbh, and it ran like shit on my 1080Ti at 1440p, I had to drop the settings down to medium to high settings to keep stable 60 fps, terrible performance.

  37. The main thing I noticed when using dlss was that jump cuts in cutscenes started to look like quick fades instead.

  38. DLSS in Control isn't actually DLSS, it's temporal reconstruction. Just a side point. I personally find the implimentation very good.

  39. Honestly, I prefer FXAA. I know, blasphemy, but hear me out. NO flickering, NO smearing artefacts, NO invisible thin things, NO first frame degrading, NO performance impact, and of course still NO jaggies. As for blur, it still blurs edges a little bit, but never by more than one pixel.

    I have a weaker PC than some people, so maybe I'm biased towards light-weight solutions. But, looking at this video, I don't think either DLSS or TAA looked particularly good. It reminds me of how video compression removes sharp edges. I guess that makes them look "cinematic" though, which seems to be a popular look.

  40. How did you justify the purchase of the (second) fastest graphics card on the market, and why do you need such a thing?

  41. 2080 Ti, the dream. The money I don't have.

  42. The reason why TAA is forced is because it is part of the RTX denoising.

  43. I dont see dust 2 here WTF… V:

Comment here